{"slug": "deepseek-v4-1-flash-occasionally-skips-thinking-even-with-thinking-enabled-max", "title": "DeepSeek-V4.1-Flash occasionally skips thinking even with thinking.enabled + reasoning_effort: max", "summary": "DeepSeek-V4.1-Flash intermittently returns responses without a thinking block even when thinking.enabled is set and reasoning_effort is max, according to an account that recommends a split-half test to determine the cause. Running a fixed panel of prompts 20 times each at a fixed temperature and comparing skip rates between the first and second halves of runs distinguishes a random dropped flag or stream from a per-query decision made on the serving or model side, which no client setting can override. Logging the raw request alongside a skipping response settles the question: if thinking is enabled in the outgoing request and no block returns, the flag arrived and was not honoured.", "body_md": "The stable per-query rate is the detail that decides this, and it is testable on your own endpoint: fix a small panel of prompts, run each one twenty times at a fixed temperature, and log whether a thinking block came back on every call. Then split each prompt’s runs in half and compare its skip rate between the halves. If a flag or a stream is being dropped at random, every prompt carries roughly the same rate in both halves. If the rate repeats per prompt — the ones that skipped early skip again late — the choice is being made per query, on the serving or model side, and no client setting will force it. Logging one skipping call’s raw request beside its response settles the last split: if the request that went out carries thinking enabled and the response still has no block, the flag arrived and was not honoured. I have a small script that runs exactly this panel and prints the split-half comparison if it would help.", "url": "https://wpnews.pro/news/deepseek-v4-1-flash-occasionally-skips-thinking-even-with-thinking-enabled-max", "canonical_source": "https://discuss.huggingface.co/t/deepseek-v4-1-flash-occasionally-skips-thinking-even-with-thinking-enabled-reasoning-effort-max/192140#post_2", "published_at": "2026-11-19 15:26:49+00:00", "updated_at": "2026-11-19 15:41:39.153617+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-infrastructure"], "entities": ["DeepSeek", "DeepSeek-V4.1-Flash"], "also_reported_by": [], "takeaway": "Run each prompt in a fixed panel 20 times at fixed temperature and compare skip rates between the first and second halves of runs: a per-prompt rate that repeats indicates a per-query serving-side decision no client flag can force, while a uniform rate points to a randomly dropped flag or stream.", "recommendation": {"advice": "Test thinking-block reliability by running each prompt 20 times at fixed temperature and comparing split-half skip rates; if the rate repeats per prompt, stop tuning client flags and log the raw request.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/deepseek-v4-1-flash-occasionally-skips-thinking-even-with-thinking-enabled-max", "markdown": "https://wpnews.pro/news/deepseek-v4-1-flash-occasionally-skips-thinking-even-with-thinking-enabled-max.md", "text": "https://wpnews.pro/news/deepseek-v4-1-flash-occasionally-skips-thinking-even-with-thinking-enabled-max.txt", "jsonld": "https://wpnews.pro/news/deepseek-v4-1-flash-occasionally-skips-thinking-even-with-thinking-enabled-max.jsonld"}}