DeepSeek Open-Sources DeepInfer Engine, Supporting Both Ascend and Blackwell Hardware Stacks

On October 20, 2026, DeepSeek open-sourced DeepInfer, the inference framework actually used internally to drive its API services, under the Apache 2.0 license. The framework supports both NVIDIA SM100 Blackwell GPUs and Huawei Ascend 950 NPUs.

Factual Recap

According to siliconangle.com, this open-sourcing marks the first time DeepSeek has publicly released its complete production-grade LLM inference infrastructure. The framework currently supports only the DeepSeek-V4.1-Flash model and loads version 1017 by default. It serves 155 trillion tokens per day, generating $38 in revenue per GPU per hour at off-peak pricing. All performance benchmarks were run on NVIDIA chips, and some features, such as deterministic output, are available only on NVIDIA hardware.

DeepInfer is based on an early vLLM fork and has integrated the FlashMLA, DeepGEMM, and DeepEP kernel libraries. The KV cache management and load balancing code has not yet been released because it is tightly coupled with internal infrastructure. The minimal efficient deployment configuration is 64 GPUs for prefill and another 64 GPUs for decode.

Mechanism Breakdown

In the DSpark speculative decoding mechanism, the draft model proposes 5 tokens each time, and the confidence head verifies only the candidates most likely to be accepted, reducing draft token verification by about 30% and delivering a 4% to 7% throughput improvement. Another technique sends only the last 128 tokens of the prompt to the decoding layer, which is said to halve prefill computation.

The Ascend 950 support stems from the September 30 adaptation of the TileLang kernel language to Huawei chips. DeepSeek plans to deploy at least 160,000 Ascend 950DT accelerators in a new data center in Inner Mongolia.

Industry Impact

The Apache 2.0 open-sourcing gives developers worldwide direct access to inference engineering capabilities previously used only inside large companies. Huawei Ascend and NVIDIA Blackwell appearing side by side in the same official documentation signals that the diversification of the AI compute supply chain has moved from discussion into engineering practice.

DeepSeek had previously open-sourced some kernel libraries, and this full framework release further lowers the barrier to large-model services. However, older architectures such as Hopper are not supported, and some core components still need to be implemented by operators themselves.

Strategic Assessment (Analysis, Not Fact)

Based on the available facts, this move may accelerate the shift away from reliance on a single hardware stack at the inference layer, but actual deployment results will still depend on the maturity of the complete ecosystem for each hardware stack. The unreleased internally coupled components show that DeepSeek is balancing openness with protecting its core competitiveness. Future model inference code will continue to be added to the same repository, a pattern similar to the evolution path of historical open-source projects.