Twenty-four hours separate two DeepSeek stories that, read together, tell you exactly where the AI wars stand.
On Monday, the US Department of Homeland Security — DHS, America’s domestic-security department — filed a public notice confirming it had ordered every US civilian federal agency to stop buying DeepSeek products and purge existing DeepSeek software from their IT systems. On Tuesday, DeepSeek answered in the most DeepSeek way imaginable: it published its inference engine, DeepInfer, on GitHub, free for anyone to inspect, copy, and use.
What did DeepSeek release?
On Tuesday, October 20, the Chinese AI lab put DeepInfer on GitHub under the Apache-2.0 license — a permissive open-source license allowing commercial use — alongside a matching tokenizer tool called deepseek-recipe. DeepSeek announced it on X, calling DeepInfer its “in-house inference framework” and saying it powers DeepSeek-V4.1-Flash “at scale with high throughput and low latency.”
What is an inference engine, anyway?
A trained AI model is a finished artifact; serving it to millions of users is a separate engineering problem, and the inference framework is the software layer that does it — splitting work across chips, batching requests so hardware stays busy, and keeping response times low.
DeepInfer bundles the usual ingredients: custom kernels, mixture-of-experts expert parallelism, speculative (draft-token) decoding, structured output, and asynchronous scheduling. It is built mostly on the open-source vLLM engine, with ideas borrowed from SGLang and TensorRT-LLM — a telling detail: even as Washington bans the company’s software, the company’s software is partly built on Western open-source tooling.
The caveat: this first release is not a complete serving stack. It handles offline inference only, serves one model (DeepSeek-V4.1-Flash), and omits a KV-cache system, prefill/decode disaggregation, and load balancing.
Why is the timing so pointed?
Because of what landed on Monday: on October 19, DHS filed a public notice in the Federal Register, confirming Exclusion and Removal Order #2026-01, issued July 24, 2026 and signed by Homeland Security Secretary Markwayne Mullin. It directs all US civilian executive-branch agencies to halt procurement of DeepSeek products and remove existing DeepSeek software — including third-party products “developed using DeepSeek implementations” — from their IT systems.
The legal authority is the Federal Acquisition Supply Chain Security Act of 2018. DHS says DeepSeek models “present multiple threat vectors,” including the risk of “transfer of sensitive data to the PRC” — the People’s Republic of China — and “actionable intelligence on US persons.” Publishing the source code one day later reads as a rebuttal in the regulators’ language: if the charge is that nobody can see inside the software, publish the software.
Does the US ban actually ban DeepSeek?
No — and the limits matter. The order covers civilian agencies only; the Pentagon and national-security systems run under separate procurement rules. It does not bind private businesses or consumers — the DeepSeek app remains available to Americans. And the 60-day exception window had already closed 28 days before the public notice appeared; China’s Foreign Ministry rejected the move the same day.
Critics have seized on the broadest phrase — products “developed using DeepSeek implementations” — since a vendor can build on DeepSeek’s outputs without shipping a single DeepSeek weight. The wariness runs both ways: Chinese apps have themselves begun blocking Chinese AI assistants.
What does the Huawei support mean?
Possibly the most strategic part: DeepInfer ships for two chip families — NVIDIA SM100 GPUs (CUDA 13, Nvidia’s programming platform) and Huawei Ascend 950 NPUs (CANN 9.2, Huawei’s equivalent), the NPUs being neural processing units purpose-built for AI workloads.
Supporting a Chinese chipmaker’s hardware alongside Nvidia’s is the point: a credible off-ramp from Nvidia for Chinese customers, and a signal to Western engineers that the code is not tied to one vendor. It lands amid Beijing’s own technology-control push — China is drafting sweeping technology export controls — and reads as a quiet declaration of chip independence.
What are the claimed performance numbers?
DeepSeek’s own figures for DeepSeek-V4.1-Flash on DeepInfer: 155 trillion tokens processed per day per cluster, a 96% input cache hit rate, roughly 400 milliseconds to first token, and about 230 tokens per second — up to 300 for coding. Those numbers come from DeepSeek and have not been independently tested: treat them as claims, not settled benchmarks.
The takeaway
Washington treats DeepSeek as a supply-chain threat; DeepSeek treats Washington’s move as a marketing problem and answers it the open-source way — transparency as strategy. Whether published code answers a national-security concern is another question: governments worry about data flows and model behavior, not just source listings.
But for the rest of the world, the release lands differently. Enterprises everywhere are asking which foreign models touch their data, and inspectable Apache-2.0 code is a genuine trust asset — while the dual-chip support quietly advances the outcome Washington most wants to slow.
