DeepSeek-V4.1-Flash

Name in the API

deepseek-ai/DeepSeek-V4.1-Flash

Base model

deepseek-ai/DeepSeek-V4.1-Flash

Weights we run

The original weights

Type

Chat

Input → output

text, image → text

Context length

1,048,576 tokens

Availability

HPC

Hardware

4 × NVIDIA H100 (94 GB), one node of the Capella HPC cluster

Thinking

Default: on, with effort high (75 on the model's scale from 1 to 100).

reasoning_effort: low (= 50), high (= 75), xhigh (= 75, the same as high) and max (= 100). minimal and medium are rejected with error 400.

Numeric effort from 1 to 100: yes, but only as chat_template_kwargs: {"reasoning_effort": 30}. A number in reasoning_effort itself is rejected with error 400.

In our tests the effort changed the amount of thinking only a little.

Turn off: reasoning_effort: "none" (no effort value at all), chat_template_kwargs: {"enable_thinking": false} or chat_template_kwargs: {"thinking": false}.

Tool calling

Yes

Example

Chat request, question about an image. Set model to the name above.

Notes

• With tool calls, send the reasoning of earlier turns back in the follow-up requests, as described in the vLLM recipe.
• The named levels do not map one to one to the levels of other providers. Use the numeric effort if you need finer control.