adtestbench Higgsfield

AI news · DeepSeek · open source

DeepSeek open-sources the engine behind its own AI service

DeepInfer, the software DeepSeek uses to serve its V4.1 Flash model, is now free on GitHub and works with Nvidia Blackwell and Huawei Ascend chips.

By Alexandre S. , 05:45 UTC

DeepSeek published DeepInfer, the software that serves its AI models to the public, as free open-source code on GitHub on Tuesday 20 October 2026. Software of this kind, called an inference framework, turns a trained model into answers for millions of users at once, so it settles much of the running cost.

Retro-futurist illustration: a striped sunset over a grid horizon under a starry sky, with a robot motif standing on the horizon.
Drawn by adtestbench from “DeepInfer: High-performance, large-scale inference framework for DeepSeek models”,

DeepSeek-V4.1-Flash is the only model it serves so far. It covers Huawei’s Ascend 950 chips as well as Nvidia’s Blackwell graphics chips. According to a deployment note shipped with the code, 155 trillion tokens a day pass through one DeepSeek cluster, SiliconANGLE reports; a token is a short chunk of text. At off-peak prices, each chip earns $38 an hour.

Two pieces stay private: the code that spreads work across machines and the code that manages the model’s short-term memory. DeepSeek says both depend on its internal systems. The code carries the Apache 2.0 licence, which allows commercial use.

In a post on X, DeepSeek said the release is meant to lower compute costs and energy use, PYMNTS reports. It came a day after the US told federal civilian agencies to remove DeepSeek’s products.

For teams that host open models themselves, most of the code a large provider uses in production is now free to copy and adapt.