DeepSeek open-sources DeepInfer, its AI model inference engine

On October 20, 2026, DeepSeek open-sourced DeepInfer, an Apache 2.0-licensed production inference framework that powers its services and enables DeepSeek-V4.1-Flash.

Compiled by Sergey KostenchukPublished 2026-10-20Updated 2026-10-20

2026-10-20 Coding NVIDIA

💻 DeepSeek open-sources DeepInfer, the engine running its AI models

On October 20, 2026, DeepSeek published the code for its production inference framework, DeepInfer, on GitHub under the Apache 2.0 license. It serves the company's real-world services and powers DeepSeek-V4.1-Flash, including the open-source kernels FlashMLA, DeepGEMM, DeepEP, and TileKernels. It claims up to 155 trillion tokens per day on a cluster. It runs on NVIDIA SM100 and Huawei Ascend 950.

🌍 For the first time, a production stack has been open-sourced, not just a model: large-scale inference becomes cheaper, and an independent alternative to vLLM and SGLang, tuned for DeepSeek, emerges.

👤 Developers can deploy DeepSeek-V4.1-Flash themselves and compare the engine with familiar alternatives. Note: Apache 2.0 covers only the code — model weight rights must be checked separately.

Source 1: https://github.com/deepseek-ai/DeepInfer

Source 2: https://cryptobriefing.com/deepseek-open-sources-deepinfer-inference-framework/

Sources