Tinker lead: DeepSeek models strong but less popular for post-training than Kimi
cHHillee · x · 2026-09-27
cHHillee (Thinking Machines) says model selection ranks fairly low on Tinker's improvement list, and asks what people actually want to post-train. Anecdotally, DeepSeek models — long strong and popular — seem less chosen for post-training than Kimi. He also acknowledges Tinker still has significant scaling work ahead.
More from Models
- Has anyone verified OpenAI's claimed Navier–Stokes solution? Community asks for external checks — miniapeur · 2026-09-27
- Dev says model benchmarks are broken: real prompts run an hour+, not $1.50 tasks — pvncher · 2026-09-27
- How models fake determinants: a 3x3 trick that dies at 4x4 and a sign-counting hack — scaling01 · 2026-09-27
- Why is Opus 5.5 so good? Insiders say it's identity and coherence, not just data — RileyRalmuto · 2026-09-27
- DeepSeek 'King of Acing Benchmarks': Dev Says New Flash Model Falls Short of Hype — bindureddy · 2026-09-27
- Little Excitement Around Gemini on My Timeline — Release Fatigue? — jobergum · 2026-09-27