DeepSeek just open-sourced DeepInfer, the inference engine behind its production services, under the Apache 2.0 license. and it got 6.3K+ stars in 8 hours The release runs only DeepSeek-V4.1-Flash, on NVIDIA SM100 GPUs or Huawei Ascend 950 chips, and covers offline inference rather than a complete serving stack. It omits the parts that turn an engine into a service: a KV cache system, prefill and decode disaggregation, and load balancing across replicas. What ships is the speed layer, including DSpark, which drafts up to 5 tokens ahead and uses confidence estimates to schedule their verification. Bounded replay trims prefill by recomputing only a 140-token suffix sized to cover the sliding attention window. A companion library, deepseek-recipe, will turn Chat Completions, Responses and Messages requests into DeepSeek V4 and V4.1 prompts that match the official API’s behaviour.

New Princeton paper finds that combining specialized agents from many developers can beat any single developer's general agent, with GAIA accuracy reaching 70%. Today most agentic systems are built and controlled by 1 developer, so each team rebuilds the same basic skills. The authors argue this limits scale. They simulated populations of developers with LLM agents that build agents under a fixed budget, then compared lone systems with a multi-developer system. The multi-developer system splits each task and routes the parts to different developers' agents. On GAIA with the same total development budget, specialists that each scored at most 35% alone reached 70% together. That is 12 points above the median general agent, while on web-search-only SealQA the combined system did worse at higher cost. Build agents to fill gaps in shared systems, since an agent scoring 19% alone lifted a combined system from 40% to 53%.

Homeland Security has ordered US civilian agencies to rip out DeepSeek and every product built on it. US hosting doesn’t get you out. The order removes all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations, from civilian agency information technology, under the Federal Acquisition Supply Chain Security Act of 2018. DHS says DeepSeek’s hosted models store data on servers in China, where national security laws give Beijing wide latitude to obtain data from domestic companies. Because every prompt and response passes through that service, DHS warned the transferred data, alone or combined with public information, could produce actionable intelligence on US persons and government programs. That logic fits the hosted chatbot, but for weights running on US servers the ban leans on DHS’s warning that the models may contain malware or other vulnerabilities. The signed order predates Monday’s public notice by months, and a July memo had already told agencies to monitor traffic in and out of government websites for DeepSeek-related risks. Exceptions require a request filed 60 days in advance with a compelling justification, such as an inability to fulfil a mission-critical function, while national security systems and secure compartmented information systems sit outside the order. The heavier work now falls on federal software vendors, who will need to show which open-weight models sit inside the products agencies already run. DeepSeek is the only Chinese model named, though Moonshot AI’s Kimi and Alibaba’s Qwen face similar distillation accusations, so the next test is whether the same order reaches them.