US and China AI regulations, DeepSeek V4.1 Flash 1017 and new open models

China and the U.S. are tightening regulatory frameworks around open AI models. Beijing proposed a pre-approval system for frontier AI models before international sharing, while Washington restricted DeepSeek from federal civilian procurement. These moves reflect a broader acceleration of geopolitical controls on AI, affecting other Chinese players like Zhipu AI already under U.S. restrictions.

DeepSeek released its inference stack DeepInfer, optimized for its models and up to 90% more cost-efficient than existing open-source tools. The company also unveiled V4.1 Flash 1017, a 552B-parameter model with image inputs under the MIT license. Xiaomi launched MiMo-V2.6-Pro-RL, an open model improved via reinforcement learning, while NVIDIA updated its Nemotron lineup. Several new U.S. and Chinese models emerged, including Beam-1.0 from Reflection, a high-performance open model under Apache 2.0.

These developments highlight a dual trend of regulatory fragmentation and technical innovation, with major players open-sourcing internal tools to dominate inference costs. Open models, once seen as unrestricted, now face geopolitical and economic pressures while remaining central to competitive advancements.