DeepSeek v4 Flash thinking mode silently ignores sampling controls in API runs
OkDimension2228 · reddit · 2026-07-29
The author spent a week debugging identifier drift in a coding agent and found that DeepSeek v4 Flash was silently ignoring sampling controls in thinking mode.
- In the hosted API path, changes to temperature, topp, presencepenalty, and frequencypenalty had no effect.
- The same prompts behaved differently on a self-hosted vLLM setup, where those parameters worked normally.
- The docs say thinking mode does not support those sampling settings and will ignore them without error, with thinking enabled by default.
- reasoningeffort is also remapped internally: low/medium → high, xhigh → max, and the docs recommend temperature=1.0 and topp=1.0.
The poster says the drift problem is still unresolved, but at least the cause is no longer mysterious.
More from Models
- Google’s AI explains hallucinations, then names Anthropic as the culprit — conitzer · 2026-07-29
- Kimi K3 faces Claude Opus 5 in a one-image 3D world-building test — Prompt Engineering · 2026-07-29
- User reports Grok 4.5 math formatting failures on web and Android — flyme2mars · 2026-07-29
- OpenAI launches GPT Transcribe and GPT Live Transcribe via API — The Decoder · 2026-07-29
- Two Minute Papers says Kimi K3 looks insane, links the paper and public demo — Two Minute Papers · 2026-07-29
- Leak says Z.ai’s GLM-5.5 targets Fable 5 and Mythos-level performance — Informal-Trouble2183 · 2026-07-29