In a spirit of trying to do what my imagination wants to be possible I set off to use or recreate Strata but to load Poolside Laguna S 2.1 Q3 on my system.
My machine has an RTX 4060 Ti 16 GB on PCI 3.0, Threadripper 1950X with AVX2, and 128 GB quad-channel DDR4-2133
I stopped after hitting a successful 84-minute autonomous run in OMP. Just 4 bugs in the result found by a frontier-ish model. Hit 770.83 tokens/sec on prefill and 15.85 tokens/sec at 8k context and 11 tokens/sec over the full run at 128k context.
The link is to the main findings and attempts at improving performance. In the end the Laguna architecture didn't lend itself to the same kind of dissection as Qwen3.8 and I've taken it as far as I'm willing to go. It's all dropped into the performance and research folders.
I love competition in the open weight world and Strata has been such a boon to my own local development I was greedy for more! Hope this helps someone, or you find it interesting at least.
In a spirit of trying to do what my imagination wants to be possible I set off to use or recreate Strata but to load Poolside Laguna S 2.1 Q3 on my system.
My machine has an RTX 4060 Ti 16 GB on PCI 3.0, Threadripper 1950X with AVX2, and 128 GB quad-channel DDR4-2133
I stopped after hitting a successful 84-minute autonomous run in OMP. Just 4 bugs in the result found by a frontier-ish model. Hit 770.83 tokens/sec on prefill and 15.85 tokens/sec at 8k context and 11 tokens/sec over the full run at 128k context.
The link is to the main findings and attempts at improving performance. In the end the Laguna architecture didn't lend itself to the same kind of dissection as Qwen3.8 and I've taken it as far as I'm willing to go. It's all dropped into the performance and research folders.
I love competition in the open weight world and Strata has been such a boon to my own local development I was greedy for more! Hope this helps someone, or you find it interesting at least.