Skip to content
AI Radar by Wagner Solution
CS
Menu
important Open-source verified update

DeepInfer framework supports both Nvidia SM100 and Huawei Ascend 950 chips

confirmed by 2 independent sources updated October 21, 2026

The DeepInfer framework for DeepSeek V4.1 Flash supports offline inference on Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. DeepSeek released its previously internal software under Apache License 2.0.

The DeepInfer framework supports offline inference for DeepSeek V4.1 Flash on Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. The report specifies the supported hardware for running the model on your own infrastructure.

DeepSeek released this previously internal inference framework under Apache License 2.0. The framework is based on vLLM and includes an optimized compute kernel, distributed processing across multiple chips and techniques for faster text generation. The company also released deepseek-recipe, a tool intended to replicate token processing in the official API.

According to DeepSeek, the release is intended to help reduce compute costs and energy consumption. The author of the Interconnects article expects greater efficiency compared with vLLM and SGLang, but this is the author's estimate without supporting comparative measurements.

What changed

Why it matters

Operators running DeepSeek V4.1 Flash can use the released software that the company used internally on supported hardware. This gives them another option for optimizing their own infrastructure; they must verify actual savings on their own workloads.

What was added since the original report

Verified updates

  1. New verified information

    DeepInfer supports offline inference on Nvidia SM100 GPUs.; DeepInfer supports offline inference on Huawei Ascend 950 NPUs.

    • DeepInfer supports offline inference on Nvidia SM100 GPUs.
    • DeepInfer supports offline inference on Huawei Ascend 950 NPUs.

Two audiences, two different impacts

What this means

01

For individuals

Developers running the model themselves also get deepseek-recipe, a tool intended to help align token processing with the official API.

What to do When running the model yourself, compare token processing using deepseek-recipe with the official API.
More practical updates →
02

For a business

Companies running DeepSeek V4.1 Flash on their own infrastructure can explore potential reductions in compute costs using the DeepInfer framework. No specific savings figure has been substantiated.

Development
What to decide Test the DeepInfer framework on supported hardware with a representative business workload and compare its speed and compute resource usage with your current solution.
More business impacts →
DeepInfer DeepSeek deepseek-recipe DeepSeek v4.1 Flash Huawei Ascend 950 Nvidia SM100 V4.1-Flash vLLM

Check the original

Event sources

confirmed by 2 independent sources · 2 publishers, 2 independent. We count feeds from the same owner only once.

3
Interconnects (Nathan Lambert, AI2) community signal · first detected Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China An overview of multiple AI topics; AI Radar covers only this event. The Decoder (daily AI news) independent context Deepseek open-sources the inference engine that powers its fastest and cheapest AI model AI Business independent context DeepSeek makes DeepInfer open source to cut AI costs