Reasoning effort on DeepSeek V4.1 Flash

DeepSeek V4.1 Flash thinks before it answers, and how long it thinks is adjustable. The setting is worth knowing because it moves the model further than most upgrades do.

Three ways it is described

• The weights. The V4.1 Flash model card describes a "continuously controllable reasoning effort setting (integer 1-100)", and DeepSeek ran its instruct benchmarks at 100.
• DeepSeek's API. Its changelog for 13 August 2026 gives the V4 thinking modes three levels, low, high and max, and suggests low for simple tasks, high for everyday agent work and max for complex ones.
• Wild West API. DeepSeek V4.1 Flash Outlaw reasons by default. The OpenAI-style reasoning_effort field takes low, high or max, and none turns reasoning off.

How much it matters

Artificial Analysis Intelligence Index v4.3.2, checked 18 October 2026.

Fourteen points separate the same weights with and without reasoning, more than separates V4.1 Flash from most models a size class above it. Reasoning tokens are output tokens and are billed as output, so effort is a trade between answer quality and spend.

Setting it

from openai import OpenAI

client = OpenAI(base_url="https://wildwestapi.com/v1", api_key="sk-ww-...")

resp = client.chat.completions.create(
    model="deepseek-v4.1-flash-outlaw",
    reasoning_effort="high",  # low, high or max; "none" turns reasoning off
    messages=[{"role": "user", "content": "Plan the refactor before writing code."}],
)
print(resp.choices[0].message.content)

The reasoning comes back alongside the answer in reasoning_content, as on DeepSeek's own API, and streams before the answer does.

Which level to use

DeepSeek's own advice is a fair default: low for lookups, formatting and short edits, high for an agent's everyday steps, max for planning, debugging and anything that failed once already. Turning reasoning off is worth it for high-volume classification or extraction, where the 25-point model is enough and the shorter output is most of the saving.

FAQ

Related