DeepSeek-V4.1-Flash

Model IDs

Model ID
`deepseek-ai/DeepSeek-V4.1-Flash`
`deepseek-ai/DeepSeek-V4.1-Flash-fast`
`deepseek-ai/DeepSeek-V4.1-Flash-ultra`

All three IDs serve the same model and accept the same parameters. Each ID has its own rates and rate limits, listed on the Models and Limits pages.

Limits

Limit	Tokens
Context length	1,048,576
Max output	393,216

The context length covers the input and max_tokens together. A request whose input plus max_tokens exceeds it returns 400. A max_tokens above 393,216 is accepted, but output stops at 393,216 tokens. A request that sets no max_tokens gets a default by mode:

Mode	Default `max_tokens`
Thinking off	8,192
Thinking on	65,536
Thinking on, effort `max`	131,072

If the prompt leaves less room than that in the context length, the default is the room left. A request with images and no max_tokens can use up to the max output. These limits and defaults match DeepSeek’s official API. Long reasoning on a large input can need more than the default, so set max_tokens when you use effort max or long documents.

Thinking

Thinking is on by default at effort high, as on DeepSeek’s official API. Control it with reasoning_effort or with a DeepSeek-style thinking object.

Request	Thinking
Neither field	On, effort `high`
`"reasoning_effort": "none"`	Off
`"reasoning_effort": "high"`	On, at the effort you set
`"thinking": {"type": "disabled"}`	Off
`"thinking": {"type": "enabled"}`	On, effort `high`
`"thinking": {"type": "enabled", "effort": "max"}`	On, at the effort you set

reasoning_effort and thinking.effort accept none, minimal, low, medium, high, xhigh, max and ultra. As on DeepSeek’s API, some run as another effort:

Requested	Runs as
`none`	Thinking off
`minimal`, `low`	`low`
`medium`, `high`, `xhigh`	`high`
`max`, `ultra`	`max`

Any other value returns 400.

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["LITHOSAI_API_KEY"],
    base_url="https://api.lithosai.cloud/v1",
)

# Thinking off
response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4.1-Flash",
    messages=[{"role": "user", "content": "What is 17 * 23?"}],
    reasoning_effort="none",
)

# Thinking on, effort max
response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4.1-Flash",
    messages=[{"role": "user", "content": "What is 17 * 23?"}],
    extra_body={"thinking": {"type": "enabled", "effort": "max"}},
)
print(response.choices[0].message.reasoning_content)
print(response.choices[0].message.content)

curl https://api.lithosai.cloud/v1/chat/completions \
  -H "Authorization: Bearer $LITHOSAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/DeepSeek-V4.1-Flash",
    "messages": [{"role": "user", "content": "What is 17 * 23?"}],
    "thinking": {"type": "disabled"}
  }'

The reasoning arrives in message.reasoning_content, or in delta.reasoning_content when streaming, separate from content. usage.completion_tokens_details.reasoning_tokens counts it. Reasoning tokens are part of completion_tokens and are billed as output. max_tokens covers the reasoning and the answer together. If the reasoning uses the whole budget, the response ends with finish_reason: "length" and an empty content. Stop sequences are matched in the reasoning too, so a match there ends the response before the answer starts.

Parameters

Parameter	Notes
`max_tokens`, `max_completion_tokens`	Cover reasoning and answer together. Default by mode, see Limits.
`temperature`	0 to 2.
`top_p`	Greater than 0, up to 1.
`top_k`, `presence_penalty`, `frequency_penalty`	Accepted.
`stop`	Also matched in the reasoning.
`seed`	Accepted, but does not make output reproducible.
`stream`, `stream_options`	Every stream ends with a usage chunk, with or without `include_usage`.
`tools`, `tool_choice`	`auto`, `required`, `none` or a named function.
`response_format`	`text`, `json_object` and `json_schema`.
`reasoning_effort`, `thinking`	See Thinking.
`user`	Accepted.

With tool_choice: "none", the tools are left out of the prompt, as on DeepSeek’s API, so the model answers in plain text. These return 400:

Parameter	`error.code`	`error.message`
`n` other than `1`	`unsupported_n`	`n must be 1`
`logprobs`, even `false`	`unsupported_parameter`	the parameter `logprobs` is not supported
`top_logprobs`	`unsupported_parameter`	the parameter `top_logprobs` is not supported
`logit_bias`	`unsupported_parameter`	the parameter `logit_bias` is not supported
Fields the API does not define, such as `chat_template_kwargs`	`unknown_parameter`	`unrecognized request argument supplied: chat_template_kwargs`

{
  "error": {
    "message": "the parameter `logprobs` is not supported",
    "type": "invalid_request_error",
    "param": null,
    "code": "unsupported_parameter"
  }
}

Out-of-range values, such as temperature: 3 or top_p: 0, also return 400.

Getting Started

Models

Inference APIs

Account

DeepSeek-V4.1-Flash

Model IDs

Limits

Thinking

Parameters

​Model IDs

​Limits

​Thinking

​Parameters

Model IDs

Limits

Thinking

Parameters