Tinker adds experimental Dimuon for LoRA fine-tuning across all its models; Thinking Machines says it reached lower loss in fewer steps than AdamW in some supervised tests.

Thinking Machines’ measured validation-loss curves comparing Dimuon and AdamW across three models and LoRA ranks.

Published 15 Oct 2026 · 04:31 UTC

Thinking Machines / Tinker ↗ · Source dated 15 Oct 2026

Browse all Live updates · Read Features