DeepSeek releases V4.1 Flash 1017 weights and prompt integration toolkit

DeepSeek released V4.1 Flash 1017 model weights under the MIT License, alongside a toolkit for integrating applications with its prompt format. The model repository describes the release as a minor update and claims improvements in reasoning, coding, agent tasks, and visual benchmarks, with shorter execution times and outputs. The deepseek-recipe toolkit converts API requests into model prompts and parses outputs, including streamed responses and tool calls. The release supplies a Python reference implementation rather than a Jinja chat template. Applications must provide model inference, tool execution, and HTTP transport themselves.