deepseek-ai/deepseek-recipe

What is it?

Key features

• 01Converts Messages, Chat Completions and Responses requests into a shared Conversation format, and parses model output back into any of them.
• 02Renders DeepSeek V4 and V4.1 conversations into prompts, or encodes them into token IDs with an attached tokenizer.
• 03Supports streaming responses, thinking mode, client tool calls and Chat Completions `prefix` continuation.
• 04Handles structured output through Chat Completions `response_format` (`json_schema` or `regex`) and Responses `text.format` (`json_schema`).
• 05Accepts images as base64 or external URLs, preprocessing them with OpenCV for V4.1.
• 06Resolves `max_tokens` against the context limit, defaulting to 8192 outside thinking mode, 65536 in thinking mode and 131072 with `max` reasoning effort.
• 07Converts `logprobs` and `top_logprobs` into `InferenceOptions` while leaving probability computation to the backend.
• 08Ships with Python bindings (`deepseek_recipe`), an Axum example server, a FastAPI example server and a browser demo.

Use cases

• Wrapping a self-hosted DeepSeek inference backend so it can serve Messages, Chat Completions and Responses clients.
• Adding streaming, tool calls or structured JSON output to an existing model endpoint without rewriting each protocol.
• Preparing training or evaluation data by rendering conversations into exact model prompts.
• Counting tokens in prompts and reasoning output, or checking `max_tokens` against the model's context limit.
• Debugging prompt formatting and output parsing through the bundled browser demo.

Quick start

1. 1

   Install the Python package

   python3 -m pip install deepseek-recipe
2. 2

   Convert a Chat Completions request and render the prompt with the example code in the README

Things to know

• !The library does not run model inference, tool execution, or HTTP transport; the caller must supply them.
• !DeepSeek V4 does not accept images, so image content is replaced with a text placeholder.
• !The Rust examples require a source build for image dependencies, which the README covers in a separate development guide.

Topics

FAQ

What is deepseek-recipe?

Rust libraries and Python bindings that convert LLM API formats into DeepSeek conversations and render prompts. deepseek-recipe is a set of Rust crates with Python bindings for putting a uniform layer between an inference backend and API clients that speak different protocols. Its job is protocol plumbing: take a request in Messages, Chat Completions or Responses format, turn it into a Conversation, render that conversation into a prompt (or token IDs) for DeepSeek models, then parse the model's output back into a response in the format the client asked for.

Is deepseek-recipe free to use? What license does it have?

deepseek-recipe is open source under the MIT license. Check the license terms before using it in commercial products.

Is deepseek-recipe still maintained?

Actively maintained. The last commit to deepseek-recipe was on Oct 19, 2026.

What can I use deepseek-recipe for?

Wrapping a self-hosted DeepSeek inference backend so it can serve Messages, Chat Completions and Responses clients. Adding streaming, tool calls or structured JSON output to an existing model endpoint without rewriting each protocol. Preparing training or evaluation data by rendering conversations into exact model prompts. Counting tokens in prompts and reasoning output, or checking `max_tokens` against the model's context limit. Debugging prompt formatting and output parsing through the bundled browser demo.

How do I install deepseek-recipe?

Install the Python package: python3 -m pip install deepseek-recipe Convert a Chat Completions request and render the prompt with the example code in the README

Where can I get deepseek-recipe?

The source code and installation instructions are on GitHub: https://github.com/deepseek-ai/deepseek-recipe