DeepSeek open-sources DeepInfer inference stack as DHS bans its products
DeepInfer ships under Apache 2.0 for Blackwell and Ascend 950, claimed to serve 155T tokens/day per cluster, the same day DHS ordered agencies to remove all DeepSeek products.
DeepSeek open-sourced DeepInfer yesterday — the actual inference stack behind its cheapest, fastest model (V4.1-Flash). Apache 2.0. Runs on Nvidia Blackwell (SM100) and Huawei Ascend 950. They claim one cluster serves 155 trillion tokens a day. Same day: DHS ordered every civilian federal agency to remove all DeepSeek products — including any third-party tools built on it — and banned future purchases. Reason: risk of sensitive data flowing to China, plus concerns about “multiple threat vectors,” actionable intelligence on US persons/agencies, and possible malware or vulnerabilities in the models.