DeepSeek V4.1 Flash post-training details leak: synthetic data drives an RL revival
joecole · x · 2026-09-10
An unverified thread on DeepSeek V4.1 Flash's post-training highlights: vanilla SFT/RL/OPD algorithms with the focus on data synthesis; two synthesis routes (general agents and coding agents built from real repos, MCP/SWE-bench style); calibration across problems, envs and verifiers for correctness and difficulty, with the model itself trained to generate tasks; reasoning effort as a scalar with effort-wise grouping and credit assignment; elastic compute for sandboxing; and exponential length penalties tied to reasoning effort. Lots of infra detail on both pre- and post-training sides.
More from Models
- KOL defends B-tier rating for GPT-6.1 Sol, blames OpenAI's botched release naming — Angaisb_ · 2026-10-27
- php2shell exploit benchmark: only 2 of 5 frontier models can port the bug — xeophon · 2026-10-27
- Qwen Code v0.26.2-preview.0 adds session permission rules and experimental shared agent loop — qwen-code-review-bot · 2026-10-27
- GPT-6.1 Terra cuts factual error rate from 22.3% to 4.3%, an ~81% drop — haider1 · 2026-10-27
- GPT-6.1 Terra lands: early users praise speed, value and subagent performance — BenBajarin · 2026-10-27
- Leaked GPT-6.1 Terra benchmarks: cheap but 'boring', still far behind Astra and Claude — Angaisb_ · 2026-10-27