The repository describes Rust libraries and Python bindings that convert requests into a shared Conversation representation and render that conversation into prompts for DeepSeek models. This release lowers integration work at the API-format boundary without removing the serving responsibilities of the integrator.
The library converts request and response formats, renders prompts, and optionally encodes conversations, while callers provide inference, tool execution, and HTTP transport. The README documents support for streaming and complete responses, tool calls, images, reasoning content, and structured output. One concrete boundary is model-specific image handling: it says DeepSeek V4 does not accept images and conversion replaces them with an unsupported-image marker, while V4.1 has image preprocessing. The mechanism is a reusable adapter layer, not a turnkey multi-provider server.
The repository presents the project as a way to connect an inference backend to services that support multiple formats, making protocol conversion a practical concern for teams exposing DeepSeek models through existing interfaces. Teams can assess deepseek-recipe as a reusable conversion layer, not a replacement for their inference or tool runtime.
The evidence is the public repository description and README; it does not establish production adoption, performance, or compatibility with every implementation of these formats. Whether future releases expand documented format coverage or provide evidence of interoperability across real serving backends. This begins a distinct analytical record because no prior New Runtime record is supplied for this project or release.