A real control arrived
DeepSeek released V4.1 Flash 1017 on October 17. The model card calls it a minor update and reports higher scores than the prior V4.1 Flash on its listed reasoning, coding, agent, and vision benchmarks. It also publishes the model weights under the MIT license.
The more useful detail for agent builders sits in the prompt encoder. V4.1 replaces the old descriptive effort labels with a numeric value from 1 through 100. The aliases map to fixed values. Low means 50, high means 75, and max means 100. The default is high.
The setting is exact. Its effect still needs a workload.
The number becomes text
The reference encoder places a sentence at the start of a thinking-mode conversation. It tells the model the chosen effort value and says a higher value should produce more thorough reasoning. The encoder only inserts that instruction at conversation index zero.
This is a prompt contract. The value does not report accelerator use. It does not promise a fixed token count, wall time, price, or quality level. A setting of 75 is not proof that the model used 75 percent of any resource.
DeepSeek also ships tests for multi-turn prompts, tools, thinking mode, numeric effort, mid-conversation system messages, and image order. Those fixtures can verify prompt construction. They cannot tell you whether a higher setting helped your repository task.
The harness owns more than the dial
The model card points production users to deepseek-recipe. That project converts Messages, Chat Completions, and Responses requests into DeepSeek prompts. It parses model output too. The repository says inference, tool execution, and HTTP transport remain the caller's job.
That boundary matters. A coding result depends on the model, prompt encoder, agent harness, tools, repository state, stop rules, and review. DeepSeek's release page recommends a minimum harness version for its own harness. A model swap without the matching prompt path can become an accidental harness test.
Road-test four readings
Choose a fixed task set with short edits, multi-file changes, debugging, and a refusal or stop case. Run the same model revision and harness at several effort values. Save the rendered prompt for each run.
Measure task result, output tokens, wall time, and tool steps. Add cost if the provider exposes it. Review the diff against the same rubric. Repeat enough work to see whether one result was luck.
- Pin the model revision and prompt encoder revision.
- Keep the task, repository revision, tools, and stop rules fixed.
- Record the numeric effort value and rendered prompt.
- Measure the result instead of inferring it from the dial.
- Choose a default for each workload class. Do not turn 100 into a universal setting.