AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 20, 2026

This fall has brought a cluster of open-weight AI releases and regulatory moves from both Washington and Beijing. DeepSeek published two new iterations of its model, including DeepSeek-V4.1-Flash-1017, a 552B-parameter model with 16B active parameters (the "A16B" designation) that accepts image inputs. The model carries an MIT license, continuing a pattern DeepSeek has established. Separately, DeepSeek released DeepInfer, its internal inference stack, which is expected to be substantially more efficient than general-purpose open tools such as vLLM and SGLang when serving DeepSeek models. Estimates cited put DeepSeek's gross margin at up to 90% when selling tokens on that stack.

The licensing detail matters. As bespoke, use-restricted licenses become more common among open-weight releases, an MIT license removes legal friction for commercial deployments, fine-tuning and redistribution. For teams choosing a base model, that is often as decisive as benchmark position.

On the regulatory side, China's Ministry of Commerce (MOFCOM) released a policy proposal, not yet law, that would require frontier AI models and some other technologies to receive government pre-approval before being shared internationally. The document states that open-source technologies would be set up for automatic approvals, which is why some observers consider the reaction overblown. Others argue the significance lies in establishing a regulatory framework that could later be tightened, and some trusted voices suggest the proposal may be aimed more at preventing situations like the sale of Chinese startup Manus to Western companies than at frontier models themselves. It is a proposal, so the practical effect on open-weight distribution is not yet determined.

In the US, the Department of Homeland Security issued an order restricting DeepSeek as a company, and DeepSeek's models as a technology platform, from serving as a vendor in federal civilian procurement. How the order treats fine-tuned derivatives or models trained on DeepSeek outputs is unclear. Nvidia's Nemotron line, for instance, has leaned heavily on DeepSeek models in its training recipes, which raises the question of whether such downstream work would be caught. The action was widely anticipated and strengthens the case for a domestic open-model ecosystem. Other Chinese developers are already constrained: Zhipu AI, maker of GLM, sits on the Commerce Department's Entity List, which limits its access to American technology.

In the picks section, Xiaomi's MiMo-V2.6-Pro-RL reuses the same base as MiMo V2.5 but spent far more compute on reinforcement learning. It ships with an in-depth technical report and some of the training environments, which is unusual at that scale. The resulting V2.6-Pro is described as the top open model on Artificial Analysis's ranking at the time of writing, and the scaled RL runs came with a live training dashboard. Cloudflare's clef is a fine-tune of a decision model, a category introduced by startup Typesafe: general classification models that emit few tokens together with their probabilities, which has driven quick adoption. Nvidia's Nemotron-3.5-Super-VL-120B-A12B-NVFP4 is the mid-sized update to the Nemotron family, strong for its size and runnable on a high-end laptop or RTX Spark, released with most of its data and a technical report. Kimi K3.1 is rumored to be imminent, with related metadata appearing in public libraries.

My reading is that the DeepInfer release may be the most consequential item for practitioners. Inference efficiency is the quiet lever on both token prices and margins, and leading labs have long kept these tools private; a public, model-specialized stack lets serving providers cut costs without changing hardware. The likely trade-off is that such a stack is tuned for DeepSeek architectures and may not transfer cleanly to other model families, so vLLM and SGLang probably retain their role as general-purpose servers. On the policy side, an Entity List restriction and a procurement ban operate differently: the former constrains a company's access to US technology, the latter bars it as a federal vendor. A procurement exclusion is narrower in reach but signals direction, and downstream ambiguity around derivative models is the part most likely to create compliance work. The MOFCOM proposal, if enacted as drafted, would leave open-source sharing on an automatic-approval path, but the framework itself is the point.

Why it matters: developers building on open weights now have a strong, permissively licensed DeepSeek option and, for the first time from a leading lab, a public inference stack that could lower serving costs across providers. Product teams selling into US federal civilian markets should treat DeepSeek-derived components as excluded, and should watch whether the ban extends to fine-tunes and models trained on DeepSeek outputs, which is inference on my part rather than stated policy. Teams sourcing Chinese open models should also track the MOFCOM proposal, since a framework that starts with automatic approvals for open source could later be adjusted.

Read at Interconnects

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Western open models strike back