Thinking Machines releases Dimuon, a Muon variant that beats AdamW for LoRA fine-tuning
clarejtbirch · x · 2026-10-15
Thinking Machines (Tinker API) released Dimuon, a variant of the Muon optimizer built for LoRA fine-tuning. In their SFT tests it often reaches lower loss in fewer steps than AdamW. The team is still experimenting with it; users can try it on Tinker today and share feedback. As John Schulman has put it, this is one way to "spend compute to save our customers compute".
Related event: Thinking Machines Releases Dimuon Optimizer for LoRA Fine-Tuning(4 posts)→
More from Models
- StepFun open-sources Step 5 Preview: 600B-param MoE, 27B active, 1M context — teortaxesTex · 2026-10-15
- Third screenshot in the comparison: wanting art to exist isn't wanting to make it — keenanisalive · 2026-10-15
- Fable's take on automating art: like driving a runner to the finish line — keenanisalive · 2026-10-15
- Astra's answer on automating art: the creative process is where the meaning lives — keenanisalive · 2026-10-15
- 9 AI Models Summarize Edited Huckleberry Finn; Only Claude Followed Both Times — GaryMarcus · 2026-10-15
- Alexandr Wang hails new Muse model's writing and orchestration, naming and pricing panned — alexandr_wang · 2026-10-15