Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China

What it claims

Issue #25 of Nathan Lambert's Interconnects "Latest open artifacts" (October 2026) argues that US and Chinese controls on open-weight models are starting to take shape. It names DeepSeek's release of its internal inference stack, DeepInfer, as one of the biggest open-source AI artifacts in months. The authors say this is the first time a leading lab has published the more efficient inference tooling that top labs keep internally. They cite estimates that DeepSeek earns up to 90% gross margin selling tokens on DeepInfer. They expect it to be "likely far more efficient" than vLLM or SGLang for DeepSeek models, and expect inference providers to adopt it quickly. The result would be lower prices for DeepSeek models across providers, or higher margins. The post gives no benchmark for DeepInfer.

Policy detail the summary omits

The MOFCOM proposal is not law. Its text says open-source technologies will get automatic approval, and the authors call the reaction "overblown." Unnamed sources they trust think it targets deals like the sale of Manus to Western companies, not frontier models.

The DHS order covers DeepSeek as a company and its models as a platform in federal civilian procurement. The authors say it is unclear whether the order reaches finetunes or models trained on DeepSeek outputs. Their example is NVIDIA's Nemotron models, which rely heavily on DeepSeek models in their training recipes. Zhipu AI (GLM) is already on the BIS Entity List. Kimi K3.1 is rumored to be imminent, with metadata appearing in public libraries.

Evidence

• MiMo-V2.6-Pro-RL (Xiaomi) uses the MiMo V2.5 bases with much more RL compute. It ships with a tech report, some of the RL environments and a live training dashboard. The authors note that releasing environments is rare at this scale. V2.6-Pro is the top open model on Artificial Analysis's ranking as of writing.
• DeepSeek-V4.1-Flash-1017 is one of two new iterations at 552B-A16B, and it accepts image inputs. A price-versus-performance chart on DeepSWE, an agentic software engineering benchmark, shows DeepSeek dominating. The text gives no figures from the chart.
• Beam-1.0-501B-A23B (Reflection, Apache 2.0) is described as the strongest permissively licensed open American model. It is not the smartest open model overall. On token efficiency it is second only to GPT-6-Astra, which the authors read as a sign of a mature post-training stack.

How to use it

• Nemotron-3.5-Super-VL-120B-A12B-NVFP4 runs locally on "a beefy laptop or RTX Spark." NVIDIA releases most of the training data and a tech report.
• Clef (Cloudflare) is a decision model: a general classifier that outputs a few tokens with probabilities. The format comes from the startup Typesafe.
• Other releases listed:
• Ternary-Bonsai-2-27B-gguf (prism-ml) is a ternary Qwen3.8-27B for on-device use.
• Ornith-1.5-397B is a set of Qwen3.5 and Gemma 4 finetunes popular with local users.
• Ling-3.0-flash-VL-fp4 (inclusionAI, Ant Group) is 124B-A5.5B.
• Muse-Spark-Open-816B-A42B is Meta's open Muse.

Limits

The fetched text is truncated partway through the General Purpose models list. Licences are stated only for DeepSeek (MIT) and Beam (Apache 2.0). Teams that build on Nemotron or other DeepSeek-derived models for US federal civilian customers should check the DHS order's scope before relying on them.