Cloud providers serving DeepSeek models on Huawei's Ascend 950 can now use the same serving code DeepSeek runs in production, though only for one model and without its caching layer.
DeepSeek has released DeepInfer, the software it uses to run its models for paying customers, as open source under the permissive Apache 2.0 license. The Hangzhou company announced the release on October 20 through its WeChat and X accounts, describing DeepInfer in its GitHub repository as the high-performance inference framework that powers its large-scale production services. The first version runs only DeepSeek-V4.1-Flash, on Nvidia’s newest Blackwell data-center GPUs (the SM100 architecture) and on Huawei’s Ascend 950 chips.
DeepSeek also published operating figures it has rarely disclosed, in a deployment document in the same repository. DeepSeek says a single production cluster serves 155 trillion tokens a day, generating revenue of US$38 per GPU-hour at off-peak prices. As Yicai’s report noted, that is a revenue figure, not a margin, and DeepSeek gave no cost data.
The release is partial. DeepInfer handles offline inference and the core functions needed to serve at scale, but it leaves out three parts of a full commercial service: a system for managing the cache of previous computations, the ability to split the processing of prompts and answers across separate machines, and load balancing across server replicas.
DeepSeek reversed its 2025 position on releasing the engine
In 2025, DeepSeek said it would not publish its inference engine. A note in its open-infrastructure repository explained that the engine was an early fork of the vLLM project, heavily customized and "tightly coupled with our internal infrastructure", and that its small team lacked the capacity to maintain a large open-source project. The New Stack observed that DeepSeek has not explained how it resolved those obstacles.
DeepInfer bundles low-level software components DeepSeek released earlier (FlashMLA, DeepGEMM and DeepEP) with speculative decoding, a technique that drafts several tokens ahead and verifies them in one pass, plus asynchronous scheduling and structured outputs such as JSON. The Next Web reported that it is built mainly on vLLM, with input from SGLang and Nvidia’s TensorRT-LLM, the open-source engines it now competes with. A companion library, deepseek-recipe, reproduces how the official API turns requests into model inputs.
DeepSeek framed the release in public-good terms. It said it was sharing its production stack to help the community lower compute costs, reduce energy use and make AI more affordable, according to The New Stack. IT Home reported that future DeepSeek models will also be supported in DeepInfer.
Huawei’s Ascend gets equal footing, with gaps
The hardware list is the most consequential detail. On Nvidia, DeepInfer supports only the SM100 architecture; older generations, including Hopper, are excluded. On Huawei, it targets the Ascend 950 through Huawei’s CANN software platform. Two features, batch-invariant inference and deterministic sampling, which make outputs repeatable, work only on Nvidia hardware.
DeepSeek’s published benchmarks use Nvidia chips alone, with four SM100 GPUs for prompt processing and 64 for generating answers. The README shows the results only as charts, and DeepSeek published no Ascend performance figures, so how the engine performs on Huawei’s chips remains an open question.
The release extends a run of DeepSeek work on Huawei’s platform. On September 30, the South China Morning Post, which Alibaba owns, reported that DeepSeek had released six open-source modules for Ascend chips, including a version of its TileLang programming tool that supports the Ascend 950. The New Stack also cited Bloomberg’s September report that DeepSeek planned to deploy at least 160,000 Ascend 950DT chips at a new data center in Inner Mongolia.
The release lands as Washington moves against DeepSeek
DeepInfer appeared the same day the US Department of Homeland Security published an order in the Federal Register requiring federal civilian agencies to remove DeepSeek products, including third-party products built using DeepSeek technology, BleepingComputer reported. Agencies have 60 days to request exceptions. The order does not cover private or consumer use.
DeepInfer follows DeepSeek’s October 17 update to V4.1-Flash and the Search and Fetch APIs EastFrontier covered. The next markers to watch are Ascend benchmark figures from DeepSeek or outside testers, and whether DeepSeek publishes the missing caching and disaggregation components that would let other providers run a full DeepSeek-grade service on Huawei hardware.