DeepSeek Open-Sources DeepInfer as the US and China Prepare Controls on Open Models

Open models are entering a phase where the rules are no longer set by benchmarks alone, but by governments as well. In the week leading up to 20 October 2026, three events landed at once: MOFCOM proposed a pre-approval system for the international transfer of frontier models, DHS cut DeepSeek off from federal civilian procurement, and DeepSeek itself open-sourced DeepInfer, its internal inference stack. Florian Brand and Nathan Lambert report this in issue 25 of the Interconnects open artifacts roundup.

Regulatory moves in Washington and Beijing

China's Ministry of Commerce has so far released only a draft policy, not a law. It would require frontier models and a number of other technologies to get state approval before they can be transferred abroad. The roundup's authors think the reaction has been overblown: the document says open technologies would receive automatic approval. Geopolitical analysts see it as a framework that could later be used to regulate China's entire AI industry. Another reading is that the proposal targets deals like the sale of the startup Manus to Western buyers, not the models themselves.

In the US, DHS issued an order under which DeepSeek as a company, and its models as a technology platform, cannot be a supplier for federal civilian procurement. The question of finetunes and models trained on DeepSeek outputs remains open. For example, Nvidia's Nemotron training recipes have long drawn on DeepSeek models, and it is unclear whether they fall under the ban. Zhipu AI, the maker of GLM, is already on the Commerce Department/BIS Entity List, which restricts its access to American technology.

DeepInfer and token economics

The most practical event of the week is the release of DeepInfer. It is DeepSeek's internal inference stack, and for its own models it is probably noticeably more efficient than the open-source vLLM and SGLang. Estimates put gross margin at up to 90% when selling tokens on this stack. It has long been known that top labs run far more efficient internal tooling than the open community has, but this is the first time a leading lab has shared it.

Inference providers will pick up DeepInfer quickly. That means one of two things: prices for DeepSeek models will fall across the ecosystem, or providers' margins will grow. The authors call the release one of the biggest open AI artifacts of recent months, even though it is neither a model nor a dataset.

What shipped in open weights

The roundup highlights five main releases:

• MiMo-V2.6-Pro-RL (Xiaomi MiMo) — an update built on the same bases as MiMo V2.5, but with far more compute spent on RL training. The authors published a detailed technical report and some of the environments, which is rare at this scale. V2.6-Pro is currently the top open model on the Artificial Analysis ranking.
• DeepSeek-V4.1-Flash-1017 (deepseek-ai) — a new 552B-A16B size with image input support, released under the MIT license. The model dominates the quality-versus-price Pareto frontier on DeepSWE, a popular agentic coding benchmark.
• clef (Cloudflare) — a "decision" model. The startup Typesafe introduced this class of models: general-purpose classifiers that output a few tokens along with probabilities. The approach caught on quickly and has spawned many open finetunes.
• NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 (nvidia) — a long-awaited update to the family's mid-sized model. It is strong for its size and runs at home if you have a powerful laptop or an RTX Spark. As usual, Nvidia publishes almost all of its data along with a detailed technical report.
• Beam-1.0-501B-A23B (Reflection) — the strongest open American model under a permissive license, Apache 2.0. It is not the smartest open model, but it is very token-efficient: it ranks second after GPT-6-Astra, which points to mature post-training.

In the general-purpose model section:

• Ternary-Bonsai-2-27B-gguf (prism-ml) — a ternary version of Qwen3.8-27B for on-device use.
• Ornith-1.5-397B (ornith-ai) — a series of Qwen3.5 and Gemma 4 finetunes that outperforms strong base models in the local community.
• Ling-3.0-flash-VL-fp4 (inclusionAI) — a 124B-A5.5B model from Ant Ling, a team inside Ant Group that releases new versions regularly.
• Muse-Spark-Open-816B-A42B (meta-models) — an open version of Meta's Muse.

Rumor has it that Kimi K3.1 is coming soon: its metadata is already appearing in public libraries.

Practical takeaways for teams

Anyone building products on open weights should check the provenance chain of their models. If your company has federal customers in the US, DeepSeek and its derivatives belong in the risk register: the DHS order does not specify how to treat distillates and finetunes. An American alternative has appeared: Beam under Apache 2.0 and Nemotron with open data offer a fallback, although Beam is not the leader in intelligence.

Inference teams should run DeepInfer against vLLM and SGLang on their own workload profiles. If the gain holds up, the cost per token of DeepSeek-V4.1-Flash-1017 on your own hardware will drop even further. Separately, decision models like clef are worth a look: routing and classification with a few tokens costs less than calling a full LLM.