qwen 3.6 struggling against its alien hand syndrome
(all qwens have this, the new ones too)
Discussion
qwen 3.6 struggling against its alien hand syndrome
(all qwens have this, the new ones too)
@artificialmind ah. i have 64GB + RTX 3060 12GB and i run a IQ3 quantized Qwen 3.8 Flash Next (the preview of Qwen 4, coming soon) through Niko1221/Strata. i get 30 tps down, 500 tps up. this model is good enough to implement entire compilers.
qwen 3.6 runs on the 32GB UMA laptop with an 860M iGPU. with gemma 4 the only other competent enough model that runs at endurable speeds
@lritter @artificialmind Strata is so weird. I have 128GB RAM so that’s pretty beefy, but only 8GB VRAM and I can run QFN at 40 TPS with 200k context and still use my computer at the same time with noticeable but not horrible slowdown. I don’t think I’ve run any other model (apart from really small really bad ones) that fast and QFN is the best one I’ve tried.