Latest Open AI Artifacts: DeepSeek V4.1 Flash 1017, American Models, and New US and Chinese Controls

Regulation and infrastructure

Open-weight AI is becoming increasingly tied to national policy. Recent developments in both the United States and China indicate that governments are considering ways to control the distribution and use of advanced models.

China’s Ministry of Commerce, or MOFCOM, released a policy proposal that is not yet law. It would require frontier AI models and certain other technologies to receive government approval before being shared internationally. The proposal says that open-source technologies would qualify for automatic approval. Some observers nevertheless view the proposal as significant because it could establish a regulatory framework for China’s AI industry.

Other analysts have suggested that the proposal may be intended primarily to prevent situations such as the sale of the Chinese startup Manus to Western companies, rather than to regulate frontier models directly. In either case, the proposal indicates that open-weight models are not outside government influence.

In the United States, the Department of Homeland Security released an order restricting DeepSeek, both as a company and as a technology platform, from serving as a vendor in federal civilian procurement. The order’s implications for fine-tuned models or models trained on DeepSeek outputs are not yet clear. For example, NVIDIA’s Nemotron models have relied heavily on DeepSeek models in their training recipes.

Other Chinese model developers, including Zhipu AI, the company behind GLM, are already on the Commerce Department’s Bureau of Industry and Security Entity List, which limits their access to American technology. These measures add to the case for developing a domestic open-model ecosystem in the United States.

DeepSeek also released its internal inference stack, DeepInfer. The software is expected to be more efficient for DeepSeek models than existing open-source tools such as vLLM and SGLang. Estimates indicate that DeepSeek could achieve gross margins of up to 90% when selling tokens using this inference stack.

Leading AI laboratories are known to operate more efficient inference systems internally, but DeepInfer is the first such release from a leading lab. Inference companies are likely to adopt the tools, potentially reducing prices for DeepSeek models across the ecosystem or increasing margins for providers. Although DeepInfer is neither a model nor a dataset, its release is one of the more significant open-source AI artifacts of recent months.

Kimi K3.1 was also rumored to be close to release, with related metadata beginning to appear in public libraries.

Featured artifacts

MiMo-V2.6-Pro-RL

XiaomiMiMo’s MiMo-V2.6-Pro-RL updates the MiMo family while retaining the same base models as MiMo V2.5. XiaomiMiMo devoted substantially more resources to reinforcement-learning training and published an in-depth technical report, along with several of the environments used during training.

The final V2.6-Pro model from this recipe ranked as the top open model on Artificial Analysis’s ranking at the time of publication. The scaled reinforcement-learning runs were accompanied by a live training dashboard.

DeepSeek-V4.1-Flash-1017

DeepSeek-V4.1-Flash-1017 followed another recent DeepSeek release, giving the company two new model iterations with a new 552B-A16B size. The models also support image inputs.

The model attracted attention for its performance at a relatively low price. It uses the MIT license, continuing DeepSeek’s practice of releasing models under a permissive license at a time when model licenses are becoming more customized and restrictive.

On the DeepSWE software-engineering benchmark, DeepSeek models occupy the performance-price frontier. DeepSWE is an agentic software-engineering benchmark focused on coding tasks.

clef

clef from Cloudflare is a decision model. This type of model produces a general classification output using a small number of tokens and provides associated probabilities.

Decision models gained attention after Typesafe introduced them, and their applications have led to rapid adoption and several open-model fine-tunes, including Clef.

NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4

NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 is the long-awaited update to the mid-sized Nemotron model family. It is designed to provide strong performance for its size and can run locally with sufficiently powerful hardware, including a high-end laptop or RTX Spark.

NVIDIA also released most of the related data and an in-depth technical report.

Beam-1.0-501B-A23B

Beam-1.0-501B-A23B is Reflection’s long-awaited open model. It uses the permissive Apache 2.0 license and was described as the strongest open American model available under such a license at the time of publication.

Beam is not the most capable open model overall, but it is highly token-efficient. Its efficiency was second only to GPT-6-Astra in the cited comparison, indicating a mature post-training system.

Other models

General-purpose models

• Ternary-Bonsai-2-27B-gguf by prism-ml: A ternary version of Qwen3.8-27B intended for on-device deployment.
• Ornith-1.5-397B by ornith-ai: A collection of Qwen3.5 and Gemma 4 fine-tunes that gained attention in the local AI community for performance above their respective baseline models.
• Ling-3.0-flash-VL-fp4 by inclusionAI: A 124B-A5.5B model from Ant Ling, part of Ant Group.
• Muse-Spark-Open-816B-A42B by meta-models: Meta’s open version of Muse.