Blogger revisits DeepSeek R1's long-CoT lesson: reasoning budget drives accuracy from 71% to 86.7%
karminski3 · x · 2026-09-14
- Blogger karminski3 revisits a key finding from the DeepSeek R1 paper: with long chain-of-thought, Pass@1 jumped from 71.0% to 86.7%.
- R1's long reasoning traces were distilled into smaller Qwen models (1.5B/7B/14B); at runtime, giving them a generous thinking budget makes accuracy scale roughly linearly with reasoning length.
- He uses this to push back on the claim that max reasoning mode isn't better — budget beyond task complexity may waste tokens, but under-budgeting clearly hurts.
- He also recalls last year's wave of people installing deepseek-r1-distilled-qwen-7b via ollama as "DeepSeek", arguing the community seems to have forgotten the lesson.
Related event: Debate Erupts Over Whether DeepSeek Thinking Intensity Affects Performance(4 posts)→
More from Models
- Users report GPT quality drop; an old-school prompt warm-up trick to regain compliance — NickPassig · 2026-10-31
- Gemini 4 Carbon vs Opus 5.5 hands-on: 3x cheaper at $4.10 and twice as fast — vitaliychiley · 2026-10-31
- 4B model Spark-X2.6 scores 94.6/100 on community benchmark, beating 27B-35B models — mesolimbic-bliss · 2026-10-31
- opencode: $100 of usage gets you DeepSeek V4.1 Flash — intellectronica · 2026-10-31
- Researcher Finds 600+ Bugs in LLM Kernel Code; Claude Blocks Safety Analysis — pmddomingos · 2026-10-31
- Signal65: Snowflake's CoCo agent finishes workflow 4.6x faster than manual — ryanshrout · 2026-10-31