deepseek-recipe ≠ DeepInfer engine

Is / is-not

Protocol surface

• Formats: Messages, Chat Completions, Responses — streaming and complete
• Prompts: DeepSeek V4 / V4.1 conversation rendering + tokenizer encoding
• Thinking mode, reasoning effort, temperature, top_p, output token limits
• Output parsing: reasoning, tool calls, JSON object, stop sequences
• Structured output: JSON Schema / regex (caller decides if backend enforces constrained decoding)
• Images (V4.1): base64 or URLs via image component; V4 replaces images with [Unsupported Image]

max_tokens defaults

From the recipe README: after encoding, ConversationRequest::resolve_max_tokens supplies defaults when the request sets none:

Python install

Python 3.10+:

python3 -m pip install deepseek-recipe

from deepseek_recipe import ChatCompletionRequest, ConversionOptions, DeepseekV41Encoding

request = ChatCompletionRequest({
    "model": "DeepSeek-V4.1-Flash",
    "messages": [{"role": "user", "content": "Hello"}],
})
converted = request.convert(ConversionOptions())
rendered = DeepseekV41Encoding().render_conversation(converted.conversation)
print(rendered.prompt)

Rust install

cargo add deepseek-recipe@0.2 deepseek-recipe-encoding@0.2 serde_json@1

See upstream development guide for source builds and image dependencies. Demo: cargo run -p encoding-decoding-demo --locked → http://127.0.0.1:7778.

Glue notes

• Recipe alone never downloads DeepInfer weights or starts an engine
• Wire recipe in front of DeepInfer offline, another backend, or the official API
• Agents / harness → dshbase.com (not recipe, not this site’s install path)
• Example servers in-repo: server-rs (Axum + mock inference), server-py (FastAPI + mock inference)

API triad (Messages · Chat Completions · Responses)

From the recipe README and production explainers: one Conversation format is shared; adapters map Messages / Chat Completions / Responses in and out. Special-token injection is blocked by encoding control tokens by ID while ordinary text that looks like tags stays literal. Streaming uses a state machine across UTF-8 / DSML / tool-argument chunk boundaries.