DeepSeek published DeepSeek-V4.1-Flash on Hugging Face on September 10, with an MIT license covering both the repository and model weights. The architecture accepts text and interleaved image inputs, producing text outputs across a context window extending up to 1 million tokens. DeepSeek reports a global key-value (KV) cache footprint of 890 bytes per token, roughly one-quarter that of DeepSeek-V4-Flash. This company-reported memory figure does not establish end-to-end serving speed or cost.
DeepSeek altered the model's tool-call tags compared to V4, meaning teams must update their prompt formatting. The repository omits a standard Jinja chat template; DeepSeek supplies a Python encoder and the deepseek-recipe Rust and Python toolkit for the format. The encoder supports mid-conversation system messages and configurable reasoning effort, accepting an integer from 1 to 100 with a reference default of 75.
For deployment, the provided inference code is a readable reference implementation, not a production serving engine. Running the reference implementation requires weight conversion, and callers remain responsible for inference hosting, tool execution, and HTTP routing.