Open models advance as US and China weigh tighter controls

Open-weight AI development is advancing alongside new regulatory pressure in the US and China. China’s Ministry of Commerce has proposed requiring government pre-approval before frontier models and certain other technologies can be shared internationally. The proposal is not law and provides for automatic approvals for open-source technologies. The US Department of Homeland Security has restricted DeepSeek and its models from federal civilian procurement, leaving uncertainty over how the restrictions apply to fine-tuned models or systems trained using DeepSeek outputs.

DeepSeek also publicly released DeepInfer, its internal inference stack. The tools could offer greater efficiency than existing open-source alternatives for its models, potentially lowering serving prices or increasing providers’ margins. Its latest Flash model supports image inputs and retains the MIT license, combining competitive performance with low pricing.

Other releases show growing breadth across the ecosystem. Xiaomi’s latest MiMo model leads Artificial Analysis’s open-model ranking, supported by expanded reinforcement learning, a detailed technical report and released training environments. Reflection’s Beam is described as the strongest American open model available under a permissive license, with notable token efficiency. Nvidia updated its mid-sized Nemotron vision-language model and released extensive data and technical documentation, while Cloudflare’s Clef joins a growing class of decision models designed to produce concise classification outputs with probabilities.