Users are glad to see Moore Threads finetuning 27B models like MusaCoder on MTT S5000 chips for kernel engineering because it shows practical RL hardware feedback loops and useful agentic results.
Based on 2 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@teortaxesTex This is exactly the kind of real-world RL + hardware feedback loop I love to see. Practical agentic workflows need models that can actually run and improve on specific chips.
@teortaxesTex nice -- glad to see more and more finetune for 27B models doing useful work I got some good results tuning QWEN for writing https://x.com/ivan_bezdomny/status/2075619939780874440?s=20
Woah Moore Threads post-trains Qwen3.6-27B for kernel engineering, getting a really good model… good RL details… but what's more salient, they test it on a MUSA port of KernelBench! And they post-trained it on their MTT S5000 chips! Chinese ecosystem grows larger https://x.com/teortaxesTex/status/2075506797918478412/photo/1 https://twitter.com/elliotarledge/status/2075485083914654029
Multiple companies showing training on domestic chinese hardware now. Hoping competition brings prices for training way down eventually https://twitter.com/teortaxesTex/status/2075506797918478412
The three-stage training pipeline optimized the model for the MUSA SDK.
Users are glad to see Moore Threads finetuning 27B models like MusaCoder on MTT S5000 chips for kernel engineering because it shows practical RL hardware feedback loops and useful agentic results.
Based on 2 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.