Two governments moved on open-weight AI this week, and neither move is clean. China's MOFCOM proposed a pre-approval requirement for exporting frontier AI models internationally. The proposal exempts open-source technologies via automatic approval, which undercuts the loudest alarm bells, but it creates a regulatory scaffolding that could tighten later. Separately, the U.S. Department of Homeland Security issued an order barring DeepSeek, the company and its model platform, from federal civilian procurement. The order leaves critical questions unanswered: Nvidia's Nemotron models rely heavily on DeepSeek outputs in their training recipes, and their status under this rule is unresolved. Zhipu AI, maker of GLM, is already on the Commerce/BIS Entity List.
DeepSeek then released DeepInfer, their internal inference stack, publicly on GitHub. This is the first time a leading lab has opened tooling of this kind. Estimates put DeepSeek's gross margin at up to 90% when serving tokens on this stack, which signals how far ahead proprietary inference tooling is compared to open alternatives like vLLM and SGLang. Inference providers will adopt DeepInfer fast. The downstream effect is either lower token prices across the ecosystem or higher margins for whoever runs it. This release is not a model or a dataset, but it is one of the more consequential open-source AI artifacts in recent months.
On the model side, Xiaomi's MiMo-V2.6-Pro-RL holds the top position on Artificial Analysis's open model rankings today. It reuses the MiMo V2.5 base but applies significantly more compute to reinforcement learning training. Xiaomi published a technical report, released training environments, and ran a live RL training dashboard at mimo.xiaomi.com/rl. Publishing environments at this scale is rare. Read the original for the full breakdown of regulatory language, the specific ambiguities in the DHS DeepSeek order, and what Kimi K3.1's imminent release signals about the competitive frontier.
[READ ORIGINAL →]