Service update · DeepSeek
DeepSeek open-sources DeepInfer for offline inference and updates its recipe library
- Review updated
DeepSeek announced DeepInfer on October 20, 2026 (Beijing time), opening the inference framework behind DeepSeek-V4.1-Flash. This release provides offline inference on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, with optimized kernels, distributed inference, speculative decoding, SWA bounded replay and structured outputs. It does not include the KV cache system, VE/prefill/decode disaggregation or cross-replica load balancing needed for a full serving system. The companion deepseek-recipe library handles API-aligned format conversion, token encoding and decoding, and output parsing. It was updated to version 0.2.0 on October 20, rather than released for the first time; callers still handle inference and tool execution.