Reasoning effort becomes a dial: low, high, max — and a 384K output ceiling
The 0731 release exposes reasoning_effort as three explicit levels, with a 384K output ceiling at the two upper settings and a recommended temperature of 1.0 with top_p 0.95 for agentic work. Compute per query is becoming a parameter the caller sets rather than a property the model has.
This is a quiet but real change in what an API is. For most of the last three years a model had a fixed cost profile and you chose between models. Exposing effort as a runtime dial means the same weights serve a cheap fast path and an expensive careful one, and the decision moves to whoever writes the call.
It also makes benchmark comparison harder in a way that will be underappreciated. A score is now a score at an effort level, and two labs quoting the same benchmark at different settings are not comparable. Any leaderboard that does not record the effort parameter is measuring something it has not defined.
The published temperature and top_p recommendations for agentic work are worth noting on their own. Vendors rarely commit to sampling settings in writing, because doing so invites reproduction. Publishing the weights alongside them makes the whole configuration checkable.
Digital Applied — DeepSeek V4 Flash 0731: official release and agent benchmarks → · BenchLM — DeepSeek V4 Flash 0731: benchmarks, MIT weights and pricing → · Umesh Malik — DeepSeek V4 Flash 0731 benchmarks: 13B active beats 1.6T →