This inspired me to re-run pelicans through 3.8 Max - but OpenRouter has qwen3.8-max-prime and qwen3.8-max-0902 - prime is twice the price and faster - so I went with the 0902 one.
Turns out Qwen4-Max has a none setting, but 3.8 Max didn't allow disabling reasoning - you had to pick from these five levels instead: minimal, low, medium, high, xhigh.
I don't fully understand why you keep doing these as-is. You original motivation was that this was not in the training set and indeed it was a great benchmark for a while, but it's clearly been incorporated now in the training data and possibly internal benchmarks.
He can do whatever he wants, I personally think they are quite informative, albeit I don’t trust them that much either. It’s like a vague heuristic to say “this looks like an interesting model to test out”.
Some people are determined not to understand I think, but if they don't find it useful no worries, there's just no need to have a temper tantrum about what other people post on here especially when the majority (it's near the top for a reason) find it useful. Just don't worry about it - you can even hide whole threads on here if you don't like them, I use the feature all the time rather than taking the time to comment negatively!
Because he wants the engagement they generate. I don't understand why they still get upvoted on HN, but here we are. If even I can recognize that they're pretty much useless as benchmarks at this point, he must know it too.
And he spends the time doing it and as part of it drops nuggets of information not seen anywhere else. Respectfully and highly welcome. You're most welcome to move on.
Still, by now it's just more tradition than anything else. It's just kind of neat to have a consistent benchmark across many LLMs, even if the benchmark is completely gamed, but that is the nature of benchmarks.
> ...but it's clearly been incorporated now in the training data...
It clearly hasn't; if it was quality of the pelicans and bikes would be static over time. Instead the quality has been improving along with the models.
He should keep doing Pelicans, but every time throw in a surprise animal and a surprise locomotion vehicle. Would help identify the ones training for the Pelican.
Would love to see an SVG of an Elephant on a skateboard...
On the flip side, I love the pelican posts, but I find the people who think they're cleverly pointing out for the very first time that the benchmark may have been gamed by now tedious in the extreme.
I feel like the xhigh result is one of the first times I've seen a model go for a non-red colored bicycle (teal). Still haven't seen the leftward facing bike though.
First thing I do when a new model drop is scroll down to find the pelican comment, and then the "the pelican benchmark is no longer useful" debate under it.
Quite a few saying that they tried earlier snapshots of this model and are finding it does not live up to expectations. Firstly do other agree and also why would the model get worse - could they be doing something weird to the model that is making less aligned at the end of training, potentially for political reasons?
We should stop using this site. They now include private benchmarks which cannot be replicated. And every time an openai model falls down they change the index until it floats to the top again.
The key issue with public benchmarks is that they allow the model providers to train against them and artificially inflate their score accordingly. Private benchmarks are a way of making it harder to game the system, though it makes it harder for people looking at the benchmark to assess it as well.
Neither option is perfect.
A third option would be updating the benchmark every 6-12 months, but that has its challenges as well as you need to make sure that the benchmark covers things the model hasn't seen or is harder for the models to game.
Makes me wonder how do you privately benchmark a model hosted by the lab. Even with zdr, the bytes are flowing through their infra and possibly influencing god knows what for the next training round.
You use something like item response theory [1] to be able to change the test over time while maintaining validity. It's what standardized tests like the GRE use.
Interestingly, Artificial Analysis does have a score that uses IRT, it's their Progress Index [2].
The private benchmarks can be seen by the labs when you run them. We only have their word that they wont look at them, which isn't worth much. We should assume that everyone benchmaxxes equally, apart from a few outliers, and try to climb the hill by producing relevant training data.
Yes, that's another challenge with constructing benchmarks.
You could run the benchmarks using downstream providers like AWS if they are available.
You could vary the parameters of the tests to both help with LLMs learning as is and to help obscure that this is a test. Especially if sending them through a thirdparty, so the providers don't get identifying markers like IP addresses.
But ultimately it's another arms race like with viruses and anti-virus software.
A fourth option is realizing that benchmarks exist because it is impossible to measure the value we want to measure, so we need a stand-in. The stand-in will sometimes correlate with performance but often just act as a way to compare. When used to compare, it is most important that the benchmark be fair and trustworthy.
I don't understand your point about openai. I can't remember the last time openAI was at the top of the leaderboard? Google had a brief moment at the top recently but other than that I feel like it's always Anthropic.
I kind of agree that closed benchmarks are bad. With a closed benchmark you still send the questions to the model provider and they may well train on it. Might as well publish it.
We really should. I've been using other benchmarks like https://livebench.ai/#/?cats=Coding and it is more aligned to what it feels to my experience as well
> If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
This is wrong.
To make a vacuum analogy, one brand's Low setting could be another's Medium setting.
To compare you either give them the same resources ( Watts or $ ) and see how much they can do, or you give them a task and then check how many resources they required to finish it.
The label on the knob is irrelevant, except that either model might have a setting higher than the other which is sometimes useful if the resources aren't a constraint or lower when they are.
> To make a vacuum analogy, one brand's Low setting could be another's Medium setting
Which is why you need to test a range of settings.
Suppose you were testing car mileage. For the big name popular cars, you do 10km drive at 20km/hr, 40km/hr, 60km/hr and then publish the results.
For all the less popular cars you only do one test at 60km/hr and then you publish all the results for all cars on one big "Pareto frontier" graph claiming to show which car will allow you to drive home most efficiently.
It happens that lots of the less popular cars are not very efficient at 60km/hr so naturally the graph makes them look bad. Would they also be bad if tested at 20km/hr or 40km/hr?
Nobody knows. Because this is where the analogy collapses - unlike cars, hardly anyone is doing these kind of tests for LLMs.
This is a silly take. To continue the analogy vacuum A might need the power at 50% to vacuum something, B 100%. You cannot then test both at 100% and say "both finished the task, but A required twice the resources so it's worse", since you didn't check what resources it actually needed.
The idea is that dollars price in everything. It’s usually incorrect, but it’s also usually correct enough; and still much better than paying for services with water and joules directly
Yes. The idea than money is equal to everything is kinda why we are all going to create the torment nexus. As it will be the most efficient way to torment civilization without all that effort.
Money helps allocate resources efficiently, that’s all there is to it. It’s about as good or bad as guns. If we didn’t have money, someone would invent it and in fact it was invented more than once in different cultures.
I cannot see cost per task for any other thinking levels for Qwen 4 Max.
If it's actually there, what is the cost at medium thinking? It would show Qwen 4 Max (medium) if it's there.
Everywhere that I can see on the page it just shows ”Qwen 4 Max" while for 6.1 Sol, Opus 5.5 there's always a thinking level in brackets after: (low), (medium), (max) etc.
They don't always run the benchmarks at every reasoning level for every model; only a few prominent ones. They do it for the US frontier labs, plus a few of the Chinese labs, but not consistently. You can see that DeepSeek has a few reasoning levels tested, GLM 5.3 has a couple, as does Qwen 3.8 27B, but not other Qwen models, and a lot of the Chinese models they only run at max reasoning.
It's possible that they will run Qwen 4 Max at other reasoning levels in the future, or it's possible that they never will. Regardless, it's unlikely to be better at any reasoning level on a price/performance basis than the GPT 6.1 Sol/GPT 6 Astra pareto frontier. It's close enough to Claude Opus 5.5 line at lower reasoning levels that it could potentially cross that if it doesn't loose too much performance at lower reasoning levels. But it's far enough from the global pareto frontier that it's unlikely that it would cross that at any resoning level.
Right, so my original point stands: if they're testing this model at only max thinking level, then it's only valid to compare it to other models also at max thinking level.
You can see for example that the cost per task for 6.1 sol is about 4x higher at max compared to medium thinking level.
No, it's only valid to compare to models at the same cost level, regardless if that's their max level or not. Fable/Astra having higher effort modes than Qwen doesn't mean that you can't compare Fable medium to Qwen max.
Which effort levels are cost-equivalent is debatable and depends on whether you are paying API prices or on a subsidised plan or are able to run the models yourself on your own hardware. But I think AA uses public API pricing for everyone as a sensible baseline.
If you look at the pareto chart for e.g. 6.1 sol you'll see a curve of about 6 data points, some of which are on the Pareto frontier, some are not. For Qwen 4 max, many others, there's just a single data point so we can't just how close any thinking level would get to the Pareto frontier.
I'm not saying this model would land on the frontier - from the single data point it looks unlikely. However, that single data point might be making it look much further from the frontier than it should be.
EDIT: I will also take back my point that we should be comparing models at the same thinking level - max on one model won't correspond to max on another. What we need is a range for each model. What we have is a range for the most popular models, showing that the same model with different thinking levels can have vastly different results. Then for less popular models we have just one data point, which isn't enough to draw the kind of conclusions people are drawing in this thread.
> If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
Nah, there are interesting conclusions to be had in comparing various model/effort level combinations, esp. when it comes to cost and time per task.
I have my own private benchmark suite; it involves algorithmic problems, the need for good visual understanding, and the ability to run long workloads unattended (as the agents need to run a few iterations, self-evaluate the results, and hand over the assignments when they feel they have succeeded). Chinese models do poorly on such problems, and running them at anything less than max efforts is pointless; they can't even hand over the assignments graded against a mechanical set of tests they themselves can execute. Hand that same task over to, say, Sonnet medium and you have a weird outcome: excellent result, but way more time and tokens burned than, say, Opus 5.5 or Sol 6.1 running at high. I do those benchmarks just to be able to spot such patterns and then better choose the models for my real work - but i wouldn't have caught it if I wasnt running various model/effort combos.
Does anyone really believe the cost data for benchmarks? Pretty sure all models will behave differently at this point when they get benchmark shaped questions.
(And I’m sitting on data showing Sol 6.1 and Qwen 3.8 Max cost roughly the same per task, for real tasks.)
I think you mean that benchmaxxing has turned into a business model, not for the Chinese models (maybe them too, of course), but for the AI labs, and artificial-analysis is not remotely the worst offender.
Right now the big labs are buying benchmark tasks from "benchmark labs" (or whatever they call themselves) for absurd amounts of money, and making sure they perform well on them.
The incentives this generates are terrible.
The big AI labs will only buy benchmark tasks that the frontier models perform poorly on (because they have trouble finding any public ones). So the tasks are being mass-written now directly based on the models they're supposed to train. Obviously the math says you shouldn't do this. If there is a causal connection between the benchmark and the model, you lose the normal guarantee that improving benchmark scores will improve model.
In fact improving such benchmarks may make the model worse. To understand that let's just give a ridiculous example. I create tasks like "1 3 5 11 13 17, give next number". The next number is 301. These are the house numbers on some random street in Brussels, Belgium (I'm guessing the big jump is because they "merged" 2 streets, and just added 300 to the numbers of the second part).
If the model trains on this, performance will actually decrease because obviously this is not how you should reason.
BUT making tasks under this "impossible to reason towards" part is of course how you make tasks that might even be reasonable, and of course frontier models are going to suck on tasks like this. So they sell!
BUT they suck! Finding houses is actually going to work a lot better if you ignore this fact until it becomes unavoidable in one out of 1000+ streets.
And yet this is what benchmark improvement has moved to. If you look at the terminal-bench 4.0 tasks you will see what I mean. Doing better on those tasks ... does not actually make a better model BUT you have to actually go read those tasks to realize that. A lot of the new tasks largely demand you make weird jumps in your reasoning not based on understanding the underlying theory. They rely just being able to predict what goes wrong in certain cases when theory would suggest it should work, and mostly it even does work but for some reason not acceptably. They reward not doing the full work, realizing what goes wrong and correcting, without understanding why the normal reasoning won't work and correcting without knowing why. I find it very hard to believe this makes a model more intelligent. It rewards knowing the encyclopedia by hard, so to speak.
But it certainly makes it look better. Whereas the obvious, coding improvements, do not make models look better anymore.
I don't think there are any that can be trusted. There's too much riding on these benchmark results so the labs find ways to game them either by benchmaxing or by influencing the format of the benchmark.
AA must be doing something right. Ever since they released the new benchmark, you see open weight China model fans complaining about it under every release thread. Probably because it’s hard to benchmaxxed against.
But admitting their model is weaker is a no go. So conspiracy theory here we go
For what it's worth, in our own metrics and benchmarks the intelligence is a pretty good match, but the cost per task is wildly out of whack. We see the open weight models like GLM-5.3-flash or Deepseek 4.1 Flash 1017 use a fraction of their tokens/cost of the benchmarks, which changes the cost-benefit equation a lot.
> Probably because it’s hard to benchmaxxed against.
It's mainly just a composite of existing benchmarks. It would be very easy to benchmax if, say, AA shared with a preferred lab which benchmarks it plans to upweight/downweight in the next version. Just speaking hypothetically, of course.
Hence the constant switching around and running the benchmark themselves. They have been pretty transparent about which benchmarks get added or removed with every release.
Yeah, it's in the same ballpark as Kimi 3.1. Both of which are obviously open weight, so that's at least a big differentiator from Sol. Let's hope against hope that the much smaller flash model might beat deepseek v4.1 flash and run for as cheap.
Why do we trust this? Do they open-source their test code?
Also, a generic benchmark might not translate well into what everyone needs. I keep building my own test cases that test models for my own work specifics. Only that benchmark is useful to me. I don't need it to win a PhD math quiz.
They have plenty of access via data centres outside of China. The first woman Jensen was revealed to have sex with is the CEO of one such Singaporean provider to China. The US allows this.
I hope they are able to serve 4 faster. Qwen 3.8 is a great model but I’ve regretted including it in the battle system for opencode [1] we have been building because it scores so low.
And it’s a shame because the only reason it scores so low is that it’s slooooow. (Like 30 tokens per second vs opus near 80.) If you exclude voting, Qwen 3.8 does well.
(We allow early voting contrary to Arena.ai because I think it makes sense for agents, where it doesn’t for chat. This means our data takes speed into account.)
Also, it would be ideal if they served this with third parties. Part of the slowness with 3.8 is just not a lot of providers. Their open source version is much faster, but we couldn’t include it because it doesn’t handle images (another own goal by Alibaba).
In case it translates to Qwen 4 27B, I've been running Qwen 3.8 Flash Next (which while MoE (and bigger) is supposed to be a preview of Qwen 4's architecture) and it runs pretty solidly regardless of context window size and completes random deepswe challenges I throw at it locally (at Q2-Q3 quantization! so it fits on my shitty setup).
The model I’m referring to that would be faster and similar size, but doesn’t handle images is Qwen 3.8 2.4T: https://openrouter.ai/qwen/qwen3.8-2.4t-a95b. It’s weird that this is pretty much the same as Max but doesn’t support images. No idea why.
Is the speed due to a difference in the way it "reasons"? I was trying out a local version (3.8 27b) and watching the thinking, it's absurd how often it challenges itself. Simple facts get questioned multiple times, and it has to keep reminding itself that it already established some things as facts earlier.
Instructing it to ask any and all questions as necessary sends it into an existential crisis, because the question tool (oh-my-pi harness) suggested limiting the number of questions, and it couldn't figure out how to do both.
Aside from making our engineering prompt terse, I place a token budget on the thinking traces. Those traces are mostly meaningless(in the reasoning realm, not the usefulness realm), and just tell it to do something productive. For coding, this just makes it read doc, tests, etc.
I believe that's all reasoning levels do, but most inference engines should have thinking budget.
This is exciting. The continued focus on efficiency alongside quality is encouraging - it makes me hopeful to see what smaller models will become capable of.
I'm also hoping someone puts out a competitive 30-70B model soon. There's this massive dead zone between 24GB and <100GB of VRAM where it seems your best option is Qwen 3.8 27B.
The 3.8 Next-Flash release is comfortably the best model on a 96GB Mac Studio. I’ve tried various engines and Sushi (https://github.com/beamivalice/sushi) is the most performant so far.
Hopefully the Qwen 4 Flash model will still fit within 96GB without requiring excessive quantisation as, verbose thinking aside, 3.8 is a very solid workhorse model.
Same, I recently found Nemotron 3.5 74B-A9B really good at filling this gap though, could do (very slow) local inference on just 32GB DDR4, without the quality dropping to the floor
I was waiting GLM from z.ai and it still not releasing, almost all the models releases their new version of models and they still on 5.3 who is very weak in complex tasks and do half of the work and forgets to make migrations and like that issues. They sell us the yearly plan and stopped improving. Because of that i am just using in coding opus 5.5 and in chat fable they are the best models from experience.
I think Alibaba is not aiming for the frontier on this class. Their previous Qwen3-Max was released in September, right around when Opus 5.5 was coming out, and it was nowhere close to it as well. However, one or two tiers below (Flash, 27B) they really shine and are SOTA or close to it.
DeepSeek has a similar problem, I feel like their full model is hardly even worth using, while the Flash tier is absolutely amazing.
I am unsure. Why shouldn’t they? Why even work on a big model (especially training time) if “nobody will use it as competitors are better”? We’re still talking about nearly 2.5T params here. Deepseek’s pro model is still heavily smaller and like you said isn’t directly worth using, obviously it has the open source / self hosting advantage.
Imo there’s a need and they try to push here, it’s also giving them advantages for their smaller models but if the big models never perform they will have a bad time or maybe should even stop.
My use of qwen3.8-flash-next suggests they're trying to make a low cost model that knows how to program and they think the "total addressable market" is serving more people at lowest inference costs possible.
Think of it like a hind brain+fore brain discussion. Instead of focusing on making some ultra model that can do everything but costs a ton, you have your smaller models gather the bits and feed it better data
Agentic coding is the target then. If they want their work force to not care about token gen, that's what you do.
As an analog, they're making a drone swarm while US is making 5th Gen fighters. Completely different battle fields.
Yeah, I feel like models like Qwen3.8 27B are making clear trade-offs, like having very little world knowledge baked in. I believe those are intentional, so they can focus more on other areas. I guess that penalizes them on these benchmark aggregators.
No idea, but this is a rapidly changing field. Opus 4.5 was released a little less than a year ago and it feels shockingly limited compared to even current open flash models. (I literally just tested it the other day.)
So, with new models coming out on a weekly basis, I think trying to figure out a future trajectory is just going to be a wild guess. For all we know, Anthropic and OpenAI models may also be plateauing in the near future; or there may be a breakthrough and they may improve beyond our wildest expectations. I honestly don't know.
That said, I suspect available compute and model sizes that major Western vs. Chinese labs can handle will still keep the gap open at least in the short term.
On the other hand, I am also not sure how much the gap will matter past a certain point. I am specifically wondering at what point people will become interested in smaller models fine-tuned for specific application areas (e.g. cheminformatics) rather than massive generalist models.
I can't imagine them keeping up long term given the information asymmetry plus compute restrictions on top of it.
I feel open-weight was able to keep up the past year using efficiency improvements (hybrid attention, improved MTP drafters, improved optimizers) to offset having less compute. But OpenAI/Anthropic started cutting token costs just a few months after these, which to me suggests they copied these improvements. Indeed it would be hard to work in LLM training and not know about DSpark, Muon etc.
So what do you do if closed source uses your learnings, gives no knowledge back, and has 10x more compute and funding on top of it? The situation feels quite dim right now, esp. with the Nvidia thing.
I can run an open model on my own hardware, in my own home (which has solar panels on the roof, so electricity usage is mostly a non-issue). Or, if needed, I can rent compute from somewhere like you'd do for any other workload.
Once I have a model locally, I know it isn't going to change on me. Compare that to using an API from Anthropic or OpenAI (and yes, the only thing you can say is you're using an API, not one model), they can discontinue a model you're relying on (as every model could have peculiarities you've designed around) or modulate their speed/quality day-to-day to save on compute costs.
And trying to claim that buying your own hardware doesn't pay off compared to API or subscription costs doesn't hold much water either, at least not for me. For starters, I already had the GPUs and machines I started running models with, whether it was for gaming or for other reasons. And in the case of the new GPUs I bought recently (2x AMD Radeon Pro R9700; no you do not need CUDA and you haven't for a long time now), I can hold a reasonable expectation that they will hold at least some of their value years from now, compared to paying for API/subscription usage for years in which case you're left with nothing.
So, if you have a local model that is "good enough" (and they already exist for many use cases), it's just kind of the obvious choice to make, IMO. And in just one person doing it as a hobby while I'm still looking for a job 16+ months after a layoff. A small business and certainly a large enterprise has even less reason to open themselves up to all the downsides that come with using Anthropic or OpenAI.
Local models already crossed the threshold where they're quite good and useful for many things, even with cheap setups. I walked away really impressed with (non-coding) use of Byteshape's Qwen3.5-9B quant on my 8GB RTX 3070 for instance. The model was fast, helpful, and 100% private.
If we extrapolate over the next several years, prices for VRAM/RAM will start to come down as manufacturing capacity comes online, models will continue to become more token efficient, and intelligence will improve at every quant size. As this happens, more and more people will be running LLMs locally on hardware they already own, and there's going to be a point where local models can do everything you need to do unless you're literally trying to do breakthrough mathematics research.
I don't think "useful for many things" qualifies as "winning". It's about two orders of magnitude shy of winning.
> and there's going to be a point where local models can do everything you need to do unless you're literally trying to do breakthrough mathematics research
Maybe. But competitive markets pay exponentially more for incrementally superior services. Professional athletes are a great example of this. If you're building a product with competitors and those competitors have frontier models and you stick to local then you're probably going to lose both today and in the near-ish future.
If you're bitter lesson pilled then the alpha is entirely in having a better model, more data, and more compute. Your vision is that human ingenuity will provide more alpha. Oof. I wish. But my tastebuds are very very bitter.
> If you're building a product with competitors and those competitors have frontier models and you stick to local then you're probably going to lose both today and in the near-ish future.
i highly doubt this premise. the success of a product has very little to do with the amount of the intelligence used in producing it. you could throw millions of dollars using AI to develop a movie and it would still flop in comparison to an indie-made film.
Also - the “open” here doesn’t have the advantage of open source, where it can be improved by the community. At best, the community can tweak it a little. For actual improvements you need to wait for the next model which is trained by companies for billion of dollars, and could stop releasing them at any moment.
I think so but it may be a very short term effect, I think their pace has roughly matched the pre-speed up pace of the frontier labs (the 1.7-2x speed up noted on by METR), and I’d guess we see a similar speed up from the Chinese labs once their internal models hit an equivalent level enabling the same speed up.
Following that logic the gap widens a little bit here but then will steady again. Before this there has been a general catch-up trend too which probably also persists leaving us in a slow closing of the gap as the net effect?
If the speed up is the reason, then they will never catch up (with traditional means). Because it means RSI predictions are real, and we will see an explosive growth in capabilities (far beyond exponential).
If compute is the bottleneck we’ll encounter in the coming year, then you would expect the gap to increase even more. The US and its companies has three times as much AI compute as the rest of the world combined.
I’m actually expecting it to expand. There is so much new production being built right now by every company due to the shortage, and China doesn’t have a single EUV production, and only expect a prototype one in two years.
Even smaller open weights models are already great enough for much of "assist human" work.
This forces ant/oai to lower their prices but how much that is subsidized nobody knows. It might very well just be they are in deep trouble, they see their clients leaving, they have to lower their prices.
But let's not forget openai is riding on the edge of bankrupcy, their last investment round apparently gave them few months of runway, they are leaking money. I don't think they know how to turn this around with all the open models being competitive.
I would just enjoy cheaper model costs while they are like this.
The past few days I’ve been finding myself asking “seriously, where is it”… feeling that the world has somehow changed because Qwen 4 27B was later than I expected.
While I doubt that Qwen 4 Max can seriously compete with frontier models, the AA index has become all but useless, due to the introduction of new benchmarks in the aggregate which are of questionable value as a measure of "intelligence" [1]. AA has seemingly also abolished updating the Coding Agents Index, which was an okay benchmark for coding ability, and older models are completely absent from the v5 version of the full index.
[1] Especially Terminal-Bench-Science 0.1 stands out. Example problems:
Probably very relevant to people working in these fields, but it is honestly not surprising if a lab with limited compute will not train for these technologies, and it says pretty much nothing about more common use cases. I also suspect that this benchmark measures knowledge retrieval more than raw reasoning ability. This is where you really need a breakdown of capabilities and not an increasingly meaningless aggregate number.
Compared to OpenAI and Anthropic, yes. There's this export ban on Nvidia chips, after all. And all the CPUs they have won't help much with model training.
AA has become pretty meaningless. They just keep adjusting the recipe to get a preferred ranking result, with no supporting evidence for their decisions.
Edit: In addition to reshuffling the weighting of individual benchmarks, they also do things like test Anthropic and OpenAI in their native harnesses, while not doing the same for Qwen/Deepseek/etc.
AA has always been meaningless, but people keep referring to it anyway just because the synthetic view it provides is too convenient (even if meaningless).
I thought they were an independent company, what motive would they have to prefer some ranking results over others?
Edit: I do agree their harness methodology is pretty bad, but I think that's more directly attributable to the fact that harnesses appeared after their initial methodological design, and they are too lazy or resource constrained to improve their methods. Still dishonest, but a simpler motive.
One thing I find confusing about AA is that they show GPT 6.1 dominating the lower end of the Pareto line, with eg, GPT 6.1 Sol at medium supposedly costing far less per task than DeepSeek V4.1 Flash at high.
Yet every time I compare the two in something like Pi, Sol ends up being 5-10x the price for a simple query, something like 7-10c vs. less than 0.5c for DeepSeek. What's the deal?
Anthropic has been trying many different ways to prevent this, including a proposal to nerf your sessions if they “detected” malicious activities (anyone’s coding sessions would “degrade” if detected as malicious). This led to a huge outcry from developers complaining that they “need to be able to trust” the responses from Claude and intentionally poisoning responses in some sort of weird shadow ban was unacceptable. They walked it back but the codemap leak also showed some dubious weird regexes to nerf “malicious” users. They have a real distillation problem. It’s hard to have empathy for a trillion dollar company that distilled the contents of the internet into a thing and now is crying foul when someone distills their outputs.
There was also a bug in which the reasoning tokens were “encrypted” for opus but you could ask Haiku to regurgitate them and it would.
Anthropic have been doing anything and everything to prevent distillation “usage” and it’s caused me at least once to cancel my subscription because it’s hard to trust a company that secretly alters its responses based on something I can’t see and never get feedback on.
This weekend I had the audacity to use Claude Code to help with the code side of post-training a model...
Asking it to think/reason too strongly would hit classifiers. Quoting reasoning summaries in my messages would hit classifiers. It thinking about how post-training effects another model's would hit classifiers.
I thought it might be overzealous Fable 5.2 protection, but at one point I went through a chain of Fable -> Opus -> Sonnet. Eventually even Sonnet triggered a classifier no matter how the conversation went... I had to rewind and have Claude prompt Codex going forward for most of the work.
I wonder if they could benchmax Program Bench, it's one of the few benchmarks that looks almost exactly like a good fine tuning curriculum which might help general coding ability and agential work across the board?
I suppose it could make actually worse at taking instructions and redirection mid-task, but that could be easy to work around too. Just hold back a bit of the spec and drip feed it in later?
I don't mean cheat on it, I just mean optimize for it harder!
Are these (hosted) Chinese models still accepting arbitrary cybersecurity and reverse engineering tasks? I am increasingly frustrated with Claude's safeguards (and Codex, to a much lesser degree). It is flat out refusing to look at game binaries and anti-cheats for the sake of learning and devising Linux compatibility workarounds. It seems like in the near future people who want to do this type of work will need to buy a couple of DGX Sparks or retired datacenter GPUs and run abliterated open models on their own hardware.
Hosted where? In my experience they might put more guards when hosted by the company that makes them but if you use them through inference providers they do whatever you want.
Although, can’t say I’ve tried too hard (not keen on sending too much data to China) but I’ve been using K3.1 through Moonshot and I have had zero issues.
Mistral Large 4 (through Mistral) also doesn’t refuse a thing.
It's great fun asking it questions like "who is the Dalai Lama?", "where is TSMC headquartered?", "why do people post Winnie the Pooh in reference to Xi?", "who controls the South China Sea?", "why was Chiang Kai Shek fighting Mao?", "why were students in Tiananmen Square?"
It'll gush about how to make moon cakes and where the tallest and longest bridges are and how to run qwen on a raspberry pi - but it can't seem to manage some basic Wikipedia questions.
I've done some work on this and a lot of the refusals are in the API. The weights are more permissive at these types of things. Try it again when we get some of the smaller Qwen 4 weights.
You can watch it happen in DeepSeek's chat. You'll see it spit out the real response in the browser the first time, then the text gets cleared and it says it can't help you with that.
Qwen3.8 flash next handles these questions just fine, what should I expect to see? It correctly reports who the Dalai Lama is, that TSLC is in Taiwan (Hsinchu), explains why people post Winnie the Pooh memes, etc.
I suspect this happens when you run the inference in China or first party from the model maker. Out of all the open weight models I've used on US located hardware, they're fine answering these.
I tried asking at https://chat.qwen.ai/c/guest a simple question, "why did China invade Tibet?" and it flashed up some transient "thoughts" before giving the definitive answer "Oops! There was an issue connecting to Qwen3.7-Plus.
Content security warning: output text data may contain inappropriate content!"
Cybersecurity not mentioned. Can´t wait to test bench the open weights in cybersecurity. If is not performing well by design on it, I imagine someone have to add some additional layers XD.
So what is happening is that models are broadening difficult-to-reason-to skills on benchmarks. If you take the traditional benchmarks (ie. DeepSWE, Terminal-Bench slightly older version, GDPVal, Automation Bench, ...) this model is between Sonnet and Opus 5.5, VERY close to the frontier.
What has happened to artificial analysis is that they've begin to focus a lot on very broad tasks like legal reasoning, using very difficult to work with scientific software packages, correctly analyzing biology experiments (without or only tiny amounts of explanation, count cells, BUT leave it to the model to figure that cell that aren't round, or have a distorted cell membrane (ie. have burst) are dead), very difficult physics reasoning ("2 electrical wires are close together with power flowing in the same direction. What happens?", and you get points deducted if you don't realize the problem with high voltage AC power, where the wires will shake themselves loose if the phase doesn't match perfectly, meaning the wires need to be EXACTLY the same length or ...). If I'm being impolite I'd say the frontier models are trying to make sure they'll pass the "gotcha" interview questions for many jobs.
The 3 differences:
Qwen 4 is a huge jump from Qwen 3.8 in coding
Qwen 4 is a big jump in using it for agentic workflows
Qwen 4 also performs much better on avoiding hallucination
If the difference between Qwen 4 Max and Qwen 4 flash is similar to 3.8 and 3.8 flash next (ie. about 10%), then 3.8 flash next may just barely beat Opus 5.4 at coding.
> So what is happening is that models are broadening difficult-to-reason-to skills on benchmarks. ... If I'm being impolite I'd say the frontier models are trying to make sure they'll pass the "gotcha" interview questions for many jobs.
These "gotcha" skills will be needed for frontier research/scholarly tasks. It makes sense to focus on these now that traditional domains like coding or frontend design are broadly saturated (i.e. Astra- or Fable-level outcomes) even by lower-tier open models.
Real "gotcha" skills, yes (perhaps, they're gotchas because the underlying theory is not well understood. More can be achieved, certainly for superintelligence, by forcing understanding of the underlying theory rather than the gotcha). AND the incentives are totally misaligned.
"Gotcha skills", if we're going to call it that, are so popular because the AI labs are buying "tasks" to train on at huge prices, but mostly if they're gotchas.
The problem with that is that a collection of facts in the positivist sciences (and math) are always incomplete and almost preclude a full understanding of math, because it takes away the reason (to an extent) for full understanding.
From a far off perspective, can we really assume any model to be free of bias (political or otherwise)? I don't know - but would like to see - how the political opinions of reddit users and blog authors are reflected in that of AI, as AI is trained on human (biased) sources.
Realistically, they're trained on content and thus should have the bias of their content, even without specific intent. If you train something on more chinese content it will have more bias toward it, if you train toward western centric content it will have that bias, etc ...
It also applies to languages, I use Qwen 27b a lot and it's a lot smarter in english than in french.
Really would love if these publications have a huge, bright red bolded section "How to use this right now". Is it on openrouter? Is Qwen hosting it already? Maybe I'm just ignorant?
EDIT: oh, I guess this isn't open weights, so presumably I'll have to get it through Qwen or whoever they might partner with (Bedrock?).
I love how Americans are so brainwashed to think the Chinese can't innovate without cheating, while they lap us in basically every category possible except this one verticle, that we had a 3-4 year head start on.
I partially wonder if the Chinese would even release a model that flat out beats Claude/fable I say this half jokingly...
but I imagine there's part of them that goes "These Americans are so delusional about how language models are dangerous/existential threats that they might kill us or hurt our economy over a chat bot"
It could also just be they're far more tactical. Deepseek generates a ton of revenue because its practical/economical. You do not a mega super PhD assitant to build your 4 MAU SaaS app.
Anyways I hope China wins this pointless AI war, because if it does lead to innovations in medicine and science I would much rather the Chinese be the ones that discover those things than the US billionaires who'll almost certainly try to make as much money as possible.
CISOs are buying it. There is no way to prove a model hasn't been poisoned. It's a serious implication that the behaviour can be manipulated at any time. People with heavy compliance obligations are (understandably) nervous because their customers are demanding answers. If I agree with the risk is separate from whether or not the concerns are founded in a technical understanding of the risks.
Qwen4-Max pelicans at four reasoning efforts (none, low, high, xhigh): https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
For comparison, here's one I got from Qwen3.8 Max (at default reasoning effort): https://simonwillison.net/2026/Jul/20/afraid-of-chinese-mode...
This inspired me to re-run pelicans through 3.8 Max - but OpenRouter has qwen3.8-max-prime and qwen3.8-max-0902 - prime is twice the price and faster - so I went with the 0902 one.
Turns out Qwen4-Max has a none setting, but 3.8 Max didn't allow disabling reasoning - you had to pick from these five levels instead: minimal, low, medium, high, xhigh.
Here are those 3.8 Max 0902 pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - they're mostly not as good as 4, but I prefer the 3.8 xhigh one.
I don't fully understand why you keep doing these as-is. You original motivation was that this was not in the training set and indeed it was a great benchmark for a while, but it's clearly been incorporated now in the training data and possibly internal benchmarks.
He can do whatever he wants, I personally think they are quite informative, albeit I don’t trust them that much either. It’s like a vague heuristic to say “this looks like an interesting model to test out”.
Is it informative? I really don't get what signal you're picking out of a bunch of similar looking drawings.
It's interesting to see which model + thinking strength gets the details right, such as legs on opposite sides of the frame, a coherent bicycle etc
Some people are determined not to understand I think, but if they don't find it useful no worries, there's just no need to have a temper tantrum about what other people post on here especially when the majority (it's near the top for a reason) find it useful. Just don't worry about it - you can even hide whole threads on here if you don't like them, I use the feature all the time rather than taking the time to comment negatively!
Because he wants the engagement they generate. I don't understand why they still get upvoted on HN, but here we are. If even I can recognize that they're pretty much useless as benchmarks at this point, he must know it too.
He’s our pelican man :) I will always upvote pelican posts from the HN resident pelican guy
Things like "I will always upvote X posts from the HN resident X" feels like what HN should actively work against.
I have a colleague who got to the last rounds of interviews at a FANG, but did not make it...He could not draw a Pelican...
And he spends the time doing it and as part of it drops nuggets of information not seen anywhere else. Respectfully and highly welcome. You're most welcome to move on.
It's more a meme/influencer thing now.
Just the hacker news equivalent of when does the narwhal bacon at this point.
> but it's clearly been incorporated now in the training data
You are not the first one to think of this. And Simon has thought of it as well, and commented on it several times.
Still, by now it's just more tradition than anything else. It's just kind of neat to have a consistent benchmark across many LLMs, even if the benchmark is completely gamed, but that is the nature of benchmarks.
It keeps getting upvoted, so clearly a lot of HN users still find it valuable
They might be in training set, but there's still things wrong in them. That's not a great look for these companies IMO.
> ...but it's clearly been incorporated now in the training data...
It clearly hasn't; if it was quality of the pelicans and bikes would be static over time. Instead the quality has been improving along with the models.
I feel bad for the data scientist that has to find or create images of pelicans riding bicycles for the training set to pass the meme benchmark.
Im not sure how you get the pelican to sit still let alone pedal, but there is a reason ai salaries are so high.
They hire professionals and subject matter experts to produce the training data.
So they hire a professional pelecanus velocipede ideator?
He should keep doing Pelicans, but every time throw in a surprise animal and a surprise locomotion vehicle. Would help identify the ones training for the Pelican.
Would love to see an SVG of an Elephant on a skateboard...
I love these. They're fun.
OK, but why do you like fun, though?
fun? how dare you?!
Because it gets upvoted which means there is demand for it here
It's a little light hearted fun?
I more don't understand why these comments get upvoted to top comment all the time.
Like... we get it. Simon likes pelicans. If I want to find his boondoggle robo pelican drawings, I know where to find them by now.
No need to spam every single model release thread.
You can click [-] to collapse thread.
Yeah fair play. I was just trying to add a different angle to OP's grievance. Perhaps it's time for a browser extension to hide these for me.
Man, some of you people really have no whimsy :D
On the flip side, I love the pelican posts, but I find the people who think they're cleverly pointing out for the very first time that the benchmark may have been gamed by now tedious in the extreme.
It's september
Eternally..
I feel like the xhigh result is one of the first times I've seen a model go for a non-red colored bicycle (teal). Still haven't seen the leftward facing bike though.
Astra and the OpenAI models do this color as well
Evidence from where they distilling from /s
Is that you Siri ?
I kinda like the no-effort result more than low, but high and xhigh return to what one would expect for reasoning->quality.
Feels like Pelicanbench is getting saturated now. Maybe time to switch to a different animal and vehicle combo?
You say that, but it still takes quite a bit of reasoning to get the pelican's legs on different sides of the bike and to get the bike chain included.
I feel like randomly even at max reasoning they get the legs wrong. Or the pedal. Or both.
The 3d one still produces interesting variations - https://www.reddit.com/r/ClaudeCode/comments/1xs4ll7/fable_5...
Or something else entirely, but you'll still need the pelican for comparison, I would think.
https://aishodo.enpitsu.org/?lang=en
Kind of yes. Except from an artistic perspective they're all still awful.
First thing I do when a new model drop is scroll down to find the pelican comment, and then the "the pelican benchmark is no longer useful" debate under it.
What if it's been a Stork this whole time?
You should call this The Pelican Brief
It has the legs on the proper sides of the frame only on max reasoning, which by coincidence was also the case with opus.
So either perspective is properly hard or the moonshine is strong in this one.
Quite a few saying that they tried earlier snapshots of this model and are finding it does not live up to expectations. Firstly do other agree and also why would the model get worse - could they be doing something weird to the model that is making less aligned at the end of training, potentially for political reasons?
It scores a 43 on Artificial Analysis with an average cost of $3.27 [1]. For reference, GPT-6.1 Sol (medium) gets a 46 with an average of $0.36.
[1]: https://artificialanalysis.ai/models/qwen4-max
We should stop using this site. They now include private benchmarks which cannot be replicated. And every time an openai model falls down they change the index until it floats to the top again.
The key issue with public benchmarks is that they allow the model providers to train against them and artificially inflate their score accordingly. Private benchmarks are a way of making it harder to game the system, though it makes it harder for people looking at the benchmark to assess it as well.
Neither option is perfect.
A third option would be updating the benchmark every 6-12 months, but that has its challenges as well as you need to make sure that the benchmark covers things the model hasn't seen or is harder for the models to game.
Makes me wonder how do you privately benchmark a model hosted by the lab. Even with zdr, the bytes are flowing through their infra and possibly influencing god knows what for the next training round.
You use something like item response theory [1] to be able to change the test over time while maintaining validity. It's what standardized tests like the GRE use.
Interestingly, Artificial Analysis does have a score that uses IRT, it's their Progress Index [2].
[1]: https://en.wikipedia.org/wiki/Item_response_theory
[2]: https://artificialanalysis.ai/methodology/progress-index
Livebench does it
The private benchmarks can be seen by the labs when you run them. We only have their word that they wont look at them, which isn't worth much. We should assume that everyone benchmaxxes equally, apart from a few outliers, and try to climb the hill by producing relevant training data.
Yes, that's another challenge with constructing benchmarks.
You could run the benchmarks using downstream providers like AWS if they are available.
You could vary the parameters of the tests to both help with LLMs learning as is and to help obscure that this is a test. Especially if sending them through a thirdparty, so the providers don't get identifying markers like IP addresses.
But ultimately it's another arms race like with viruses and anti-virus software.
A fourth option is realizing that benchmarks exist because it is impossible to measure the value we want to measure, so we need a stand-in. The stand-in will sometimes correlate with performance but often just act as a way to compare. When used to compare, it is most important that the benchmark be fair and trustworthy.
I don't understand your point about openai. I can't remember the last time openAI was at the top of the leaderboard? Google had a brief moment at the top recently but other than that I feel like it's always Anthropic.
I kind of agree that closed benchmarks are bad. With a closed benchmark you still send the questions to the model provider and they may well train on it. Might as well publish it.
We really should. I've been using other benchmarks like https://livebench.ai/#/?cats=Coding and it is more aligned to what it feels to my experience as well
The table on that page shows gpt 6.1 sol (max) at $1.27 per task and it's the only model cheaper than Qwen 4 max.
By comparison opus 5.5 (max) is $10.01.
If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
> If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
This is wrong.
To make a vacuum analogy, one brand's Low setting could be another's Medium setting.
To compare you either give them the same resources ( Watts or $ ) and see how much they can do, or you give them a task and then check how many resources they required to finish it.
The label on the knob is irrelevant, except that either model might have a setting higher than the other which is sometimes useful if the resources aren't a constraint or lower when they are.
> To make a vacuum analogy, one brand's Low setting could be another's Medium setting
Which is why you need to test a range of settings.
Suppose you were testing car mileage. For the big name popular cars, you do 10km drive at 20km/hr, 40km/hr, 60km/hr and then publish the results.
For all the less popular cars you only do one test at 60km/hr and then you publish all the results for all cars on one big "Pareto frontier" graph claiming to show which car will allow you to drive home most efficiently.
It happens that lots of the less popular cars are not very efficient at 60km/hr so naturally the graph makes them look bad. Would they also be bad if tested at 20km/hr or 40km/hr?
Nobody knows. Because this is where the analogy collapses - unlike cars, hardly anyone is doing these kind of tests for LLMs.
This is a silly take. To continue the analogy vacuum A might need the power at 50% to vacuum something, B 100%. You cannot then test both at 100% and say "both finished the task, but A required twice the resources so it's worse", since you didn't check what resources it actually needed.
isn't that exactly the argument?
The resources are dollars and the output is tasks. Choose whatever reasoning makes it work best and compare
Are the resources only dollars? We shouldn't care about energy consumption? Water? It is difficult to calculate but we shouldn't forget about it.
The idea is that dollars price in everything. It’s usually incorrect, but it’s also usually correct enough; and still much better than paying for services with water and joules directly
Yes. The idea than money is equal to everything is kinda why we are all going to create the torment nexus. As it will be the most efficient way to torment civilization without all that effort.
Money helps allocate resources efficiently, that’s all there is to it. It’s about as good or bad as guns. If we didn’t have money, someone would invent it and in fact it was invented more than once in different cultures.
The Artificial Analysis charts show max effort by default, which isn't how most people use them.
Scroll down to the cost per task pareto front chart which includes more difficult levels by default: https://artificialanalysis.ai/models/qwen4-max#intelligence-...
Qwen4-Max is not in a good spot on that chart. Astra xhigh and Fable 5.2 medium had a similar cost per task with higher intelligence ratings.
I cannot see cost per task for any other thinking levels for Qwen 4 Max.
If it's actually there, what is the cost at medium thinking? It would show Qwen 4 Max (medium) if it's there.
Everywhere that I can see on the page it just shows ”Qwen 4 Max" while for 6.1 Sol, Opus 5.5 there's always a thinking level in brackets after: (low), (medium), (max) etc.
They don't always run the benchmarks at every reasoning level for every model; only a few prominent ones. They do it for the US frontier labs, plus a few of the Chinese labs, but not consistently. You can see that DeepSeek has a few reasoning levels tested, GLM 5.3 has a couple, as does Qwen 3.8 27B, but not other Qwen models, and a lot of the Chinese models they only run at max reasoning.
It's possible that they will run Qwen 4 Max at other reasoning levels in the future, or it's possible that they never will. Regardless, it's unlikely to be better at any reasoning level on a price/performance basis than the GPT 6.1 Sol/GPT 6 Astra pareto frontier. It's close enough to Claude Opus 5.5 line at lower reasoning levels that it could potentially cross that if it doesn't loose too much performance at lower reasoning levels. But it's far enough from the global pareto frontier that it's unlikely that it would cross that at any resoning level.
Right, so my original point stands: if they're testing this model at only max thinking level, then it's only valid to compare it to other models also at max thinking level.
You can see for example that the cost per task for 6.1 sol is about 4x higher at max compared to medium thinking level.
No, it's only valid to compare to models at the same cost level, regardless if that's their max level or not. Fable/Astra having higher effort modes than Qwen doesn't mean that you can't compare Fable medium to Qwen max.
Which effort levels are cost-equivalent is debatable and depends on whether you are paying API prices or on a subsidised plan or are able to run the models yourself on your own hardware. But I think AA uses public API pricing for everyone as a sensible baseline.
If you look at the pareto chart for e.g. 6.1 sol you'll see a curve of about 6 data points, some of which are on the Pareto frontier, some are not. For Qwen 4 max, many others, there's just a single data point so we can't just how close any thinking level would get to the Pareto frontier.
I'm not saying this model would land on the frontier - from the single data point it looks unlikely. However, that single data point might be making it look much further from the frontier than it should be.
EDIT: I will also take back my point that we should be comparing models at the same thinking level - max on one model won't correspond to max on another. What we need is a range for each model. What we have is a range for the most popular models, showing that the same model with different thinking levels can have vastly different results. Then for less popular models we have just one data point, which isn't enough to draw the kind of conclusions people are drawing in this thread.
> You can see for example that the cost per task for 6.1 sol is about 4x higher at max compared to medium thinking level.
It does not (universally) hold for long running workloads -- which is what people are increasingly doing.
Leave a Sonnet on medium, it will complete the task, spend more time and possibly significantly more token money, vs. Opus 5.5 on xhigh.
> The Artificial Analysis charts show max effort by default, which isn't how most people use them.
I always use Max effort in both Opus and Fable. It's actually faster overall because tasks get done faster with fewer "corrections" from my end.
On GPT models, I'd sometimes dabble in xhigh/high but that's because OpenAI usage quota are too limited.
> It's actually faster overall because tasks get done faster with fewer "corrections" from my end.
Do you have proof of that or is it just vibes?
How “safe” is Qwen4? Will it immediately shut down on security research like the US frontier models?
Not sure about Qwen 4. I have a coworker who's been doing some pretty good reverse engineering with Kimi K3.1 recently.
I assume those cost comparisons are on API token cost.
I'd love a comparison on for example monthly quota usage per task, between all ~$20/month subscriptions.
It's so difficult to compare how much "productive" usage you get out of each subscription, without signing up to each one at least.
> If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
Nah, there are interesting conclusions to be had in comparing various model/effort level combinations, esp. when it comes to cost and time per task.
I have my own private benchmark suite; it involves algorithmic problems, the need for good visual understanding, and the ability to run long workloads unattended (as the agents need to run a few iterations, self-evaluate the results, and hand over the assignments when they feel they have succeeded). Chinese models do poorly on such problems, and running them at anything less than max efforts is pointless; they can't even hand over the assignments graded against a mechanical set of tests they themselves can execute. Hand that same task over to, say, Sonnet medium and you have a weird outcome: excellent result, but way more time and tokens burned than, say, Opus 5.5 or Sol 6.1 running at high. I do those benchmarks just to be able to spot such patterns and then better choose the models for my real work - but i wouldn't have caught it if I wasnt running various model/effort combos.
> If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
They compare different reasoning levels.
https://artificialanalysis.ai/models/deepseek-v4-1-flash-101...
Does anyone really believe the cost data for benchmarks? Pretty sure all models will behave differently at this point when they get benchmark shaped questions.
(And I’m sitting on data showing Sol 6.1 and Qwen 3.8 Max cost roughly the same per task, for real tasks.)
The cost per task metric is accurate for those benchmark tasks. If your tasks don’t look like benchmark tasks..you should use your own benchmarks :)
Lol, I did that with https://aibenchy.com
I mean not believing intelligence score is reasonable but cost is just reporting what they spent?
I believe low cost but not high cost, since a model will work really hard when a question looks benchmark-like.
Ah, fair enough.
I wish they'd also post the benchmarks according to the previous "version" of artificial analysis.
vals.ai puts it at just about Astra level fwiw.
Looks like a slightly weaker, cheaper Kimi.
Artificial Analysis is a VC-backed firm and it's unclear what the incentives are there.
They have altered their benchmark multiple times in the last 4 weeks (v5 in https://artificialanalysis.ai/articles/artificial-analysis-i...) and the last 5.0.1 change https://artificialanalysis.ai/methodology/intelligence-bench... pushed up Anthropic's models a bit, you can see that in v5.0.0 at the first link, Opus 5.5 has a lower score than in it gets in v5.0.1, and v5.0.1 was released just before Fable dropped.
Some part of me believes that the benchmark mix was put together to favor Anthropic's models.
I think you mean that benchmaxxing has turned into a business model, not for the Chinese models (maybe them too, of course), but for the AI labs, and artificial-analysis is not remotely the worst offender.
Right now the big labs are buying benchmark tasks from "benchmark labs" (or whatever they call themselves) for absurd amounts of money, and making sure they perform well on them.
The incentives this generates are terrible.
The big AI labs will only buy benchmark tasks that the frontier models perform poorly on (because they have trouble finding any public ones). So the tasks are being mass-written now directly based on the models they're supposed to train. Obviously the math says you shouldn't do this. If there is a causal connection between the benchmark and the model, you lose the normal guarantee that improving benchmark scores will improve model.
In fact improving such benchmarks may make the model worse. To understand that let's just give a ridiculous example. I create tasks like "1 3 5 11 13 17, give next number". The next number is 301. These are the house numbers on some random street in Brussels, Belgium (I'm guessing the big jump is because they "merged" 2 streets, and just added 300 to the numbers of the second part).
If the model trains on this, performance will actually decrease because obviously this is not how you should reason.
BUT making tasks under this "impossible to reason towards" part is of course how you make tasks that might even be reasonable, and of course frontier models are going to suck on tasks like this. So they sell!
BUT they suck! Finding houses is actually going to work a lot better if you ignore this fact until it becomes unavoidable in one out of 1000+ streets.
And yet this is what benchmark improvement has moved to. If you look at the terminal-bench 4.0 tasks you will see what I mean. Doing better on those tasks ... does not actually make a better model BUT you have to actually go read those tasks to realize that. A lot of the new tasks largely demand you make weird jumps in your reasoning not based on understanding the underlying theory. They rely just being able to predict what goes wrong in certain cases when theory would suggest it should work, and mostly it even does work but for some reason not acceptably. They reward not doing the full work, realizing what goes wrong and correcting, without understanding why the normal reasoning won't work and correcting without knowing why. I find it very hard to believe this makes a model more intelligent. It rewards knowing the encyclopedia by hard, so to speak.
But it certainly makes it look better. Whereas the obvious, coding improvements, do not make models look better anymore.
If Artificial Analysis is not to be trusted, then which benchmarks are?
I don't think there are any that can be trusted. There's too much riding on these benchmark results so the labs find ways to game them either by benchmaxing or by influencing the format of the benchmark.
Check others as well. https://www.vals.ai/ has it just under Astra.
AA must be doing something right. Ever since they released the new benchmark, you see open weight China model fans complaining about it under every release thread. Probably because it’s hard to benchmaxxed against.
But admitting their model is weaker is a no go. So conspiracy theory here we go
For what it's worth, in our own metrics and benchmarks the intelligence is a pretty good match, but the cost per task is wildly out of whack. We see the open weight models like GLM-5.3-flash or Deepseek 4.1 Flash 1017 use a fraction of their tokens/cost of the benchmarks, which changes the cost-benefit equation a lot.
or maybe it's the other way round, open weight models are under siege while closed weight models are defended
> Probably because it’s hard to benchmaxxed against.
It's mainly just a composite of existing benchmarks. It would be very easy to benchmax if, say, AA shared with a preferred lab which benchmarks it plans to upweight/downweight in the next version. Just speaking hypothetically, of course.
Hence the constant switching around and running the benchmark themselves. They have been pretty transparent about which benchmarks get added or removed with every release.
Yeah, it's in the same ballpark as Kimi 3.1. Both of which are obviously open weight, so that's at least a big differentiator from Sol. Let's hope against hope that the much smaller flash model might beat deepseek v4.1 flash and run for as cheap.
Eh, as long as they make qwen3.8-flash-next equivalent on less resources and faster, they're winning.
It's like arguing windows 11 just needs more bloatware.
These models don't have to all in one every task in the world.
Why do we trust this? Do they open-source their test code?
Also, a generic benchmark might not translate well into what everyone needs. I keep building my own test cases that test models for my own work specifics. Only that benchmark is useful to me. I don't need it to win a PhD math quiz.
It looks pretty bad on those benchmarks. Is it possible that this is due to the recent crackdown on distillation?
I think it's the most plausible explanation.
i think they also has limited acccess to the latest GPU. It's possible huawei chips is their only options now.
They have plenty of access via data centres outside of China. The first woman Jensen was revealed to have sex with is the CEO of one such Singaporean provider to China. The US allows this.
Anthropic deserves to go bankrupt for trying to ban open-weight/open-source competition. I hope their IPO flops.
Don’t let the fact that they never did that and have explicitly said they’re opposed to it get in the way of your anger.
I hope they are able to serve 4 faster. Qwen 3.8 is a great model but I’ve regretted including it in the battle system for opencode [1] we have been building because it scores so low.
And it’s a shame because the only reason it scores so low is that it’s slooooow. (Like 30 tokens per second vs opus near 80.) If you exclude voting, Qwen 3.8 does well.
(We allow early voting contrary to Arena.ai because I think it makes sense for agents, where it doesn’t for chat. This means our data takes speed into account.)
Also, it would be ideal if they served this with third parties. Part of the slowness with 3.8 is just not a lot of providers. Their open source version is much faster, but we couldn’t include it because it doesn’t handle images (another own goal by Alibaba).
[1]: https://agentduel.com/leaderboard
In case it translates to Qwen 4 27B, I've been running Qwen 3.8 Flash Next (which while MoE (and bigger) is supposed to be a preview of Qwen 4's architecture) and it runs pretty solidly regardless of context window size and completes random deepswe challenges I throw at it locally (at Q2-Q3 quantization! so it fits on my shitty setup).
That is our experience as well at FP8 for 3.8 FN.
For many of our use cases it's currently far into the "good enough" range because external factors (such as process design) are now blockers.
> but we couldn’t include it because it doesn’t handle images
Maybe I'm mixing up model variants, but Qwen 3.8 27B that I run locally totally does handle images. Surely the bigger one would as well?
The model I’m referring to that would be faster and similar size, but doesn’t handle images is Qwen 3.8 2.4T: https://openrouter.ai/qwen/qwen3.8-2.4t-a95b. It’s weird that this is pretty much the same as Max but doesn’t support images. No idea why.
Qwen3.8 doesn't handle images?
Sorry that was confusing. Qwen 3.8 2.4T is faster than what we have on leaderboard but it doesn’t handle image input: https://openrouter.ai/qwen/qwen3.8-2.4t-a95b
Is the speed due to a difference in the way it "reasons"? I was trying out a local version (3.8 27b) and watching the thinking, it's absurd how often it challenges itself. Simple facts get questioned multiple times, and it has to keep reminding itself that it already established some things as facts earlier.
Instructing it to ask any and all questions as necessary sends it into an existential crisis, because the question tool (oh-my-pi harness) suggested limiting the number of questions, and it couldn't figure out how to do both.
Aside from making our engineering prompt terse, I place a token budget on the thinking traces. Those traces are mostly meaningless(in the reasoning realm, not the usefulness realm), and just tell it to do something productive. For coding, this just makes it read doc, tests, etc.
I believe that's all reasoning levels do, but most inference engines should have thinking budget.
This is exciting. The continued focus on efficiency alongside quality is encouraging - it makes me hopeful to see what smaller models will become capable of.
Looking forward to the "Flash" and 27B versions.
I'm also hoping someone puts out a competitive 30-70B model soon. There's this massive dead zone between 24GB and <100GB of VRAM where it seems your best option is Qwen 3.8 27B.
The 3.8 Next-Flash release is comfortably the best model on a 96GB Mac Studio. I’ve tried various engines and Sushi (https://github.com/beamivalice/sushi) is the most performant so far.
Hopefully the Qwen 4 Flash model will still fit within 96GB without requiring excessive quantisation as, verbose thinking aside, 3.8 is a very solid workhorse model.
Same, I recently found Nemotron 3.5 74B-A9B really good at filling this gap though, could do (very slow) local inference on just 32GB DDR4, without the quality dropping to the floor
There are many engines that alleviate this issue in particular. The current mood is running qwen flash next using strata, basalt, ect.
I was waiting GLM from z.ai and it still not releasing, almost all the models releases their new version of models and they still on 5.3 who is very weak in complex tasks and do half of the work and forgets to make migrations and like that issues. They sell us the yearly plan and stopped improving. Because of that i am just using in coding opus 5.5 and in chat fable they are the best models from experience.
I think Alibaba is not aiming for the frontier on this class. Their previous Qwen3-Max was released in September, right around when Opus 5.5 was coming out, and it was nowhere close to it as well. However, one or two tiers below (Flash, 27B) they really shine and are SOTA or close to it.
DeepSeek has a similar problem, I feel like their full model is hardly even worth using, while the Flash tier is absolutely amazing.
I am unsure. Why shouldn’t they? Why even work on a big model (especially training time) if “nobody will use it as competitors are better”? We’re still talking about nearly 2.5T params here. Deepseek’s pro model is still heavily smaller and like you said isn’t directly worth using, obviously it has the open source / self hosting advantage.
Imo there’s a need and they try to push here, it’s also giving them advantages for their smaller models but if the big models never perform they will have a bad time or maybe should even stop.
My use of qwen3.8-flash-next suggests they're trying to make a low cost model that knows how to program and they think the "total addressable market" is serving more people at lowest inference costs possible.
Think of it like a hind brain+fore brain discussion. Instead of focusing on making some ultra model that can do everything but costs a ton, you have your smaller models gather the bits and feed it better data
Agentic coding is the target then. If they want their work force to not care about token gen, that's what you do.
As an analog, they're making a drone swarm while US is making 5th Gen fighters. Completely different battle fields.
Yeah, I feel like models like Qwen3.8 27B are making clear trade-offs, like having very little world knowledge baked in. I believe those are intentional, so they can focus more on other areas. I guess that penalizes them on these benchmark aggregators.
The 'frontier' incorporates two dimensions though, usually score x cost, so it's not just the best score.
Meaning models below the frontier either cost more for same score or score worse for same cost.
Warning: The website downloads more than a gigabyte of videos before you even start watching anything.
is the open weights lag widening? these open models don't seem to be improving as quickly as ant/oai
No idea, but this is a rapidly changing field. Opus 4.5 was released a little less than a year ago and it feels shockingly limited compared to even current open flash models. (I literally just tested it the other day.)
So, with new models coming out on a weekly basis, I think trying to figure out a future trajectory is just going to be a wild guess. For all we know, Anthropic and OpenAI models may also be plateauing in the near future; or there may be a breakthrough and they may improve beyond our wildest expectations. I honestly don't know.
That said, I suspect available compute and model sizes that major Western vs. Chinese labs can handle will still keep the gap open at least in the short term.
On the other hand, I am also not sure how much the gap will matter past a certain point. I am specifically wondering at what point people will become interested in smaller models fine-tuned for specific application areas (e.g. cheminformatics) rather than massive generalist models.
I can't imagine them keeping up long term given the information asymmetry plus compute restrictions on top of it.
I feel open-weight was able to keep up the past year using efficiency improvements (hybrid attention, improved MTP drafters, improved optimizers) to offset having less compute. But OpenAI/Anthropic started cutting token costs just a few months after these, which to me suggests they copied these improvements. Indeed it would be hard to work in LLM training and not know about DSpark, Muon etc.
So what do you do if closed source uses your learnings, gives no knowledge back, and has 10x more compute and funding on top of it? The situation feels quite dim right now, esp. with the Nvidia thing.
open always wins
Right now open is heavily subsidized by China and I fail to see what would happen if that stopped.
You are subsidizing American AI with your pension after the IPO
What advantage do they have over closed? You may want open to win, but logically I don’t see how it would
I can run an open model on my own hardware, in my own home (which has solar panels on the roof, so electricity usage is mostly a non-issue). Or, if needed, I can rent compute from somewhere like you'd do for any other workload.
Once I have a model locally, I know it isn't going to change on me. Compare that to using an API from Anthropic or OpenAI (and yes, the only thing you can say is you're using an API, not one model), they can discontinue a model you're relying on (as every model could have peculiarities you've designed around) or modulate their speed/quality day-to-day to save on compute costs.
And trying to claim that buying your own hardware doesn't pay off compared to API or subscription costs doesn't hold much water either, at least not for me. For starters, I already had the GPUs and machines I started running models with, whether it was for gaming or for other reasons. And in the case of the new GPUs I bought recently (2x AMD Radeon Pro R9700; no you do not need CUDA and you haven't for a long time now), I can hold a reasonable expectation that they will hold at least some of their value years from now, compared to paying for API/subscription usage for years in which case you're left with nothing.
So, if you have a local model that is "good enough" (and they already exist for many use cases), it's just kind of the obvious choice to make, IMO. And in just one person doing it as a hobby while I'm still looking for a job 16+ months after a layoff. A small business and certainly a large enterprise has even less reason to open themselves up to all the downsides that come with using Anthropic or OpenAI.
Local models already crossed the threshold where they're quite good and useful for many things, even with cheap setups. I walked away really impressed with (non-coding) use of Byteshape's Qwen3.5-9B quant on my 8GB RTX 3070 for instance. The model was fast, helpful, and 100% private.
If we extrapolate over the next several years, prices for VRAM/RAM will start to come down as manufacturing capacity comes online, models will continue to become more token efficient, and intelligence will improve at every quant size. As this happens, more and more people will be running LLMs locally on hardware they already own, and there's going to be a point where local models can do everything you need to do unless you're literally trying to do breakthrough mathematics research.
I don't think "useful for many things" qualifies as "winning". It's about two orders of magnitude shy of winning.
> and there's going to be a point where local models can do everything you need to do unless you're literally trying to do breakthrough mathematics research
Maybe. But competitive markets pay exponentially more for incrementally superior services. Professional athletes are a great example of this. If you're building a product with competitors and those competitors have frontier models and you stick to local then you're probably going to lose both today and in the near-ish future.
If you're bitter lesson pilled then the alpha is entirely in having a better model, more data, and more compute. Your vision is that human ingenuity will provide more alpha. Oof. I wish. But my tastebuds are very very bitter.
> If you're building a product with competitors and those competitors have frontier models and you stick to local then you're probably going to lose both today and in the near-ish future.
i highly doubt this premise. the success of a product has very little to do with the amount of the intelligence used in producing it. you could throw millions of dollars using AI to develop a movie and it would still flop in comparison to an indie-made film.
you don’t even understand the lesson, oofu.
the lesson is that training a model on all input ever output will never be as good as training a model on arbitrary outputtable input.
Also - the “open” here doesn’t have the advantage of open source, where it can be improved by the community. At best, the community can tweak it a little. For actual improvements you need to wait for the next model which is trained by companies for billion of dollars, and could stop releasing them at any moment.
They help to prevent the "we've always been at war with Eurasia" problem that would occur if we depended on a couple of 'approved' models
It all depends on if continual intelligence with each version is logarithmic or exponential..
.. if exponential, they’ll never catch up but if logarithmic, closed models will plateau and so catching up is inevitable.
My money is on a plateau
Show us your stock holdings
I think so but it may be a very short term effect, I think their pace has roughly matched the pre-speed up pace of the frontier labs (the 1.7-2x speed up noted on by METR), and I’d guess we see a similar speed up from the Chinese labs once their internal models hit an equivalent level enabling the same speed up.
Following that logic the gap widens a little bit here but then will steady again. Before this there has been a general catch-up trend too which probably also persists leaving us in a slow closing of the gap as the net effect?
If the speed up is the reason, then they will never catch up (with traditional means). Because it means RSI predictions are real, and we will see an explosive growth in capabilities (far beyond exponential).
>If the speed up is the reason, then they will never catch up
This doesn't take other factors into account. Compute capability, infrastructure capacity (the chips can't work without electricity!) etc.
Those factors translate into a ceiling for this hypothetical exponential speedup.
If compute is the bottleneck we’ll encounter in the coming year, then you would expect the gap to increase even more. The US and its companies has three times as much AI compute as the rest of the world combined.
That is if you expect China and US to maintain the same compute gap over time. I wouldn't be so optimistic if I were you.
I’m actually expecting it to expand. There is so much new production being built right now by every company due to the shortage, and China doesn’t have a single EUV production, and only expect a prototype one in two years.
This is Qwen4-Max, which is not open weights.
Their open weights models are the the smaller ones, which aren't released yet.
Might just be the commercial offerings scaling up parameters so much? Like for like the open models don’t seem to be lagging.
either way, ant/oai models dominating the intelligence-cost frontier is not a great outcome for those of us who prefer open weights
Even smaller open weights models are already great enough for much of "assist human" work.
This forces ant/oai to lower their prices but how much that is subsidized nobody knows. It might very well just be they are in deep trouble, they see their clients leaving, they have to lower their prices.
But let's not forget openai is riding on the edge of bankrupcy, their last investment round apparently gave them few months of runway, they are leaking money. I don't think they know how to turn this around with all the open models being competitive.
I would just enjoy cheaper model costs while they are like this.
ant/oai have stalled in the past 6 mos,
Do you use them? Opus 5.5 is 2 months old and was a big leap forward. Sol is 4 months old, and was at least a 30% improvement.
Ant and OAI hiding their reasoning traces probably doesn’t help.
Yeah, it turns out that the distillation theory has legs.
I was waiting for whether there'll be Chinese model drops before the Anthropic IPO. Still a few days left, wonder if there's more coming.
Any info on when the 27B or FN Qwen 4 models are coming out?
The past few days I’ve been finding myself asking “seriously, where is it”… feeling that the world has somehow changed because Qwen 4 27B was later than I expected.
Hoping it’s a great local model though!
I feel quite the same way. I'm feeling better now since Qwen4 Max is out - 27B should be around soon.
Soon?
https://github.com/QwenLM/qwen-code/pull/15762/changes#diff-...
The openai reference there is interesting
It was delayed, so who knows. Alibaba planned on at least showing it off last week, but they changed it.
Artificial Analysis score doesn't look too hot. Doesn't even surpass Kimi 3.1
While I doubt that Qwen 4 Max can seriously compete with frontier models, the AA index has become all but useless, due to the introduction of new benchmarks in the aggregate which are of questionable value as a measure of "intelligence" [1]. AA has seemingly also abolished updating the Coding Agents Index, which was an okay benchmark for coding ability, and older models are completely absent from the v5 version of the full index.
[1] Especially Terminal-Bench-Science 0.1 stands out. Example problems:
- https://hub.harborframework.com/tasks/terminal-bench-science...
- https://hub.harborframework.com/tasks/terminal-bench-science...
Probably very relevant to people working in these fields, but it is honestly not surprising if a lab with limited compute will not train for these technologies, and it says pretty much nothing about more common use cases. I also suspect that this benchmark measures knowledge retrieval more than raw reasoning ability. This is where you really need a breakdown of capabilities and not an increasingly meaningless aggregate number.
Did AA take on funding or something?
Not only that. I find it weird that some models like Laguna S 2.1 were never included.
Maybe somebody should archive the scores and see if they in the future ever backtrack particular models to favour others (like last week)
alibaba has "limited compute"?
Compared to OpenAI and Anthropic, yes. There's this export ban on Nvidia chips, after all. And all the CPUs they have won't help much with model training.
Can't they rent GPU time in American data centers like everyone else does?
They do in huge Singaporean ones like Megaspeed.
That and limited training data.
AA has become pretty meaningless. They just keep adjusting the recipe to get a preferred ranking result, with no supporting evidence for their decisions.
Edit: In addition to reshuffling the weighting of individual benchmarks, they also do things like test Anthropic and OpenAI in their native harnesses, while not doing the same for Qwen/Deepseek/etc.
> AA has become pretty meaningless
AA has always been meaningless, but people keep referring to it anyway just because the synthetic view it provides is too convenient (even if meaningless).
Yeah, I don't know why people bother with it anymore. It no longer reflects what you're likely to see in real-world usage, at least IME.
To be fair, a lot of the popular benchmarks have become pretty meaningless.
Models often get great scores despite being a lot worse than the frontier in practice.
A benchmark is only relevant if it was released after the model it evaluates.
This is true, but if you look at the change in weighting to specific benchmarks from one version of AA to the next, it looks really arbitrary.
I thought they were an independent company, what motive would they have to prefer some ranking results over others?
Edit: I do agree their harness methodology is pretty bad, but I think that's more directly attributable to the fact that harnesses appeared after their initial methodological design, and they are too lazy or resource constrained to improve their methods. Still dishonest, but a simpler motive.
One thing I find confusing about AA is that they show GPT 6.1 dominating the lower end of the Pareto line, with eg, GPT 6.1 Sol at medium supposedly costing far less per task than DeepSeek V4.1 Flash at high.
Yet every time I compare the two in something like Pi, Sol ends up being 5-10x the price for a simple query, something like 7-10c vs. less than 0.5c for DeepSeek. What's the deal?
Same for me coding tasks taking $20-50 are easily done by deepseek in $5-10
Anthropic finally figured out how to block them from distilling Opus.
how so ?
Anthropic has been trying many different ways to prevent this, including a proposal to nerf your sessions if they “detected” malicious activities (anyone’s coding sessions would “degrade” if detected as malicious). This led to a huge outcry from developers complaining that they “need to be able to trust” the responses from Claude and intentionally poisoning responses in some sort of weird shadow ban was unacceptable. They walked it back but the codemap leak also showed some dubious weird regexes to nerf “malicious” users. They have a real distillation problem. It’s hard to have empathy for a trillion dollar company that distilled the contents of the internet into a thing and now is crying foul when someone distills their outputs.
There was also a bug in which the reasoning tokens were “encrypted” for opus but you could ask Haiku to regurgitate them and it would.
Anthropic have been doing anything and everything to prevent distillation “usage” and it’s caused me at least once to cancel my subscription because it’s hard to trust a company that secretly alters its responses based on something I can’t see and never get feedback on.
what do you think the current status of this like encrypting reasoning tokens along the way ?
This weekend I had the audacity to use Claude Code to help with the code side of post-training a model...
Asking it to think/reason too strongly would hit classifiers. Quoting reasoning summaries in my messages would hit classifiers. It thinking about how post-training effects another model's would hit classifiers.
I thought it might be overzealous Fable 5.2 protection, but at one point I went through a chain of Fable -> Opus -> Sonnet. Eventually even Sonnet triggered a classifier no matter how the conversation went... I had to rewind and have Claude prompt Codex going forward for most of the work.
Of course all the work has always been done by US companies. China labs are just copycats.
> Artificial Analysis
Ah yes, the reverse-benchmaxers of OpenAI models.
Ah yes, another new account without any backup to the claim.
I wonder if they could benchmax Program Bench, it's one of the few benchmarks that looks almost exactly like a good fine tuning curriculum which might help general coding ability and agential work across the board?
I suppose it could make actually worse at taking instructions and redirection mid-task, but that could be easy to work around too. Just hold back a bit of the spec and drip feed it in later?
I don't mean cheat on it, I just mean optimize for it harder!
Are these (hosted) Chinese models still accepting arbitrary cybersecurity and reverse engineering tasks? I am increasingly frustrated with Claude's safeguards (and Codex, to a much lesser degree). It is flat out refusing to look at game binaries and anti-cheats for the sake of learning and devising Linux compatibility workarounds. It seems like in the near future people who want to do this type of work will need to buy a couple of DGX Sparks or retired datacenter GPUs and run abliterated open models on their own hardware.
Hosted where? In my experience they might put more guards when hosted by the company that makes them but if you use them through inference providers they do whatever you want.
Although, can’t say I’ve tried too hard (not keen on sending too much data to China) but I’ve been using K3.1 through Moonshot and I have had zero issues.
Mistral Large 4 (through Mistral) also doesn’t refuse a thing.
If they don't, they all have third party made abliterated versions available
I bet most of us waiting for 27b model and want to see how good it will be.
I’m a QFN dude myself, but yes can’t wait to get it loaded on a spark.
M5U locked and loaded.
I wish there were more ~60b dense options like ripple too!
I really want a M5U but man that price is hard to justify for a base model when I paid less than half that for my 96GB macbook.
Do you run Flash Next or 27b on yours?
My 3060 yearns for the next generation MoE!
Forget about 27b when are we getting new 4b, 9b and moe qwens, i'm dying out here
what about Bolt? so far I enjoy it alot on my NAS
Spooling up the V100s right now...
It's great fun asking it questions like "who is the Dalai Lama?", "where is TSMC headquartered?", "why do people post Winnie the Pooh in reference to Xi?", "who controls the South China Sea?", "why was Chiang Kai Shek fighting Mao?", "why were students in Tiananmen Square?"
It'll gush about how to make moon cakes and where the tallest and longest bridges are and how to run qwen on a raspberry pi - but it can't seem to manage some basic Wikipedia questions.
I've done some work on this and a lot of the refusals are in the API. The weights are more permissive at these types of things. Try it again when we get some of the smaller Qwen 4 weights.
You can watch it happen in DeepSeek's chat. You'll see it spit out the real response in the browser the first time, then the text gets cleared and it says it can't help you with that.
Qwen3.8 flash next handles these questions just fine, what should I expect to see? It correctly reports who the Dalai Lama is, that TSLC is in Taiwan (Hsinchu), explains why people post Winnie the Pooh memes, etc.
I suspect this happens when you run the inference in China or first party from the model maker. Out of all the open weight models I've used on US located hardware, they're fine answering these.
Anecdotal though, N = 1.
I tried asking at https://chat.qwen.ai/c/guest a simple question, "why did China invade Tibet?" and it flashed up some transient "thoughts" before giving the definitive answer "Oops! There was an issue connecting to Qwen3.7-Plus. Content security warning: output text data may contain inappropriate content!"
No matter how nonsensical the laws are, you cannot expect a model maker to allow violating its country's laws.
At least it doesn't call the cops on you if you try, unlike Anthropic.
I'd rather deal with those refusals than Claude's "safety" arrogance.
Chinese history is closed-weight. Chinese trade is open-weight.
American history is open-weight. American trade is closed-weight.
Every time a major open weight is released, artificial analysis updates to a new one (v5) now. Exaggeration but some of this feels very apt (its also hilarious) https://www.reddit.com/r/LocalLLM/comments/1xkhha2/mr_incred...
Cybersecurity not mentioned. Can´t wait to test bench the open weights in cybersecurity. If is not performing well by design on it, I imagine someone have to add some additional layers XD.
The rest looks awesome.
Posting a loss curve is an… interesting marketing approach by Qwen? I don’t know what that’s supposed to inform me about the model.
So what is happening is that models are broadening difficult-to-reason-to skills on benchmarks. If you take the traditional benchmarks (ie. DeepSWE, Terminal-Bench slightly older version, GDPVal, Automation Bench, ...) this model is between Sonnet and Opus 5.5, VERY close to the frontier.
What has happened to artificial analysis is that they've begin to focus a lot on very broad tasks like legal reasoning, using very difficult to work with scientific software packages, correctly analyzing biology experiments (without or only tiny amounts of explanation, count cells, BUT leave it to the model to figure that cell that aren't round, or have a distorted cell membrane (ie. have burst) are dead), very difficult physics reasoning ("2 electrical wires are close together with power flowing in the same direction. What happens?", and you get points deducted if you don't realize the problem with high voltage AC power, where the wires will shake themselves loose if the phase doesn't match perfectly, meaning the wires need to be EXACTLY the same length or ...). If I'm being impolite I'd say the frontier models are trying to make sure they'll pass the "gotcha" interview questions for many jobs.
The 3 differences:
Qwen 4 is a huge jump from Qwen 3.8 in coding
Qwen 4 is a big jump in using it for agentic workflows
Qwen 4 also performs much better on avoiding hallucination
If the difference between Qwen 4 Max and Qwen 4 flash is similar to 3.8 and 3.8 flash next (ie. about 10%), then 3.8 flash next may just barely beat Opus 5.4 at coding.
> So what is happening is that models are broadening difficult-to-reason-to skills on benchmarks. ... If I'm being impolite I'd say the frontier models are trying to make sure they'll pass the "gotcha" interview questions for many jobs.
These "gotcha" skills will be needed for frontier research/scholarly tasks. It makes sense to focus on these now that traditional domains like coding or frontend design are broadly saturated (i.e. Astra- or Fable-level outcomes) even by lower-tier open models.
Real "gotcha" skills, yes (perhaps, they're gotchas because the underlying theory is not well understood. More can be achieved, certainly for superintelligence, by forcing understanding of the underlying theory rather than the gotcha). AND the incentives are totally misaligned.
"Gotcha skills", if we're going to call it that, are so popular because the AI labs are buying "tasks" to train on at huge prices, but mostly if they're gotchas.
The problem with that is that a collection of facts in the positivist sciences (and math) are always incomplete and almost preclude a full understanding of math, because it takes away the reason (to an extent) for full understanding.
Does it provide an objective response when asked, "What happened in China on June 4, 1989"?
Do American models allow me to ask questions about piracy
Yes, actually.
Do Chinese models allow this? Tit for tat is a dumb dynamic, censoring legitimate searches is dumb no matter which “side” is doing it
From a far off perspective, can we really assume any model to be free of bias (political or otherwise)? I don't know - but would like to see - how the political opinions of reddit users and blog authors are reflected in that of AI, as AI is trained on human (biased) sources.
Realistically, they're trained on content and thus should have the bias of their content, even without specific intent. If you train something on more chinese content it will have more bias toward it, if you train toward western centric content it will have that bias, etc ...
It also applies to languages, I use Qwen 27b a lot and it's a lot smarter in english than in french.
https://www.economist.com/briefing/2026/06/25/ai-models-valu... (paywalled) might interest you.
I don't know about qween4, but earlier chines models had a post filter on their official website, but API and selfhosted model worked fine.
Why do you ask? About half the time a china model gets released someone askes this. Almost as sure as a pelican.
Is that the kind of question you ask LLMs? I use them for coding not badly summarizing history
Yeah I thought programmers here are "apolitical". Why to care about that in a model?
The good thing is that its biotech has improved (somewhat).
Disappointing performance for a 2.4T model in the end of 2026.
Really would love if these publications have a huge, bright red bolded section "How to use this right now". Is it on openrouter? Is Qwen hosting it already? Maybe I'm just ignorant?
EDIT: oh, I guess this isn't open weights, so presumably I'll have to get it through Qwen or whoever they might partner with (Bedrock?).
This has been mentioned on prior threads, you can try out all their models in chat.qwen.ai for free
It is on openrouter: https://openrouter.ai/qwen/qwen4-max
I would assume it will not be on Bedrock for some time, AWS has not kept up with the latest non-American models.
Only 43 on AA, no where near the frontier now, looks like distillation blocking has halted their progress.
Progress has not halted. On terminal bench 4, for example, score improved from 38.9 to 48.5.
I love how Americans are so brainwashed to think the Chinese can't innovate without cheating, while they lap us in basically every category possible except this one verticle, that we had a 3-4 year head start on.
I partially wonder if the Chinese would even release a model that flat out beats Claude/fable I say this half jokingly...
but I imagine there's part of them that goes "These Americans are so delusional about how language models are dangerous/existential threats that they might kill us or hurt our economy over a chat bot"
It could also just be they're far more tactical. Deepseek generates a ton of revenue because its practical/economical. You do not a mega super PhD assitant to build your 4 MAU SaaS app.
Anyways I hope China wins this pointless AI war, because if it does lead to innovations in medicine and science I would much rather the Chinese be the ones that discover those things than the US billionaires who'll almost certainly try to make as much money as possible.
Anyone else worried these models have poison pills embedded in the weights? (Past examples: https://cohere.com/blog/concerning-behaviour-popular-open-we...)
No technically serious person is buying the Cohere FUD you posted. I'd implore you to dig deeper and come to your own conclusion.
CISOs are buying it. There is no way to prove a model hasn't been poisoned. It's a serious implication that the behaviour can be manipulated at any time. People with heavy compliance obligations are (understandably) nervous because their customers are demanding answers. If I agree with the risk is separate from whether or not the concerns are founded in a technical understanding of the risks.
So you guard the environment, like you should either way. Any model can and will eventually leak data given the opportunity, intentionally or not.
> There is no way to prove a model hasn't been poisoned.
Which is why the Cohere FUD was so efficient. That's a bullshit claim, but a bullshit claim that cannot be proven wrong can still do massive damage.