✳ the wire · analysis
DeepSeek open-sources DeepInfer, the engine behind V4.1 Flash, built for NVIDIA Blackwell and Huawei Ascend 950
An email the second DeepSeek’s next model answers on a public API, plus every frontier launch · no card
DeepSeek released DeepInfer, the inference framework it says serves DeepSeek-V4.1-Flash at scale, as open source on October 20 under the Apache 2.0 license. It also open-sourced deepseek-recipe, the token encoding and decoding component it says follows the same processing as the official DeepSeek API, so community deployments behave like the API.
WHAT IS IN IT
Per the repository's README, V4.1 Flash is the one supported model so far. The engine uses DeepSeek's own FlashMLA, DeepGEMM and DeepEP kernels, speculative decoding with draft tokens, a bounded-replay mode that cuts prefill work for sliding-window attention, and batch-invariant inference and deterministic sampling on NVIDIA hardware. DeepSeek says inference code for its future models will land in the same repository.
THE HARDWARE
Two targets: NVIDIA SM100 GPUs, the Blackwell generation, and Huawei Ascend 950 NPUs, which need Huawei's CANN 9.2. That lands a day after The Information reported, citing two people familiar with the matter, that DeepSeek is working closely with Huawei to shape the design of future AI chips.
WHAT WE DO NOT KNOW
Which model the repository adds next, and how DeepInfer compares with vLLM or SGLang on the same hardware. DeepSeek's README publishes its own throughput curves, measured on 4 SM100 GPUs for prefill and 64 for decode, but no head-to-head.
Source: DeepSeek on X (launch thread, read directly) · github.com/deepseek-ai/DeepInfer README (read directly) · @Mr_Salio and @scaling01 on X ↗ · DeepSeek V4.1 Flash tracker · the bench index
