China's Ministry of Commerce added 20 Japanese entities to its export control list today, blocking Chinese exporters from sending dual use items to four Japanese defense research institutes and nine Mitsubishi defense subsidiaries, Caixin reported. A second batch of 20 Japanese firms was placed on a separate watch list. A ministry spokesperson said the move was a response to Japan's "remilitarization" and nuclear ambitions, adding that it targets only a small number of entities and would not disrupt broader China-Japan trade. China's dual use export regime, first sharpened on rare earths and chip making inputs, is now being applied to a U.S. treaty ally, expanding the ministry's retaliation toolkit against allied suppliers feeding U.S. AI and defense supply chains.
Read at Caixin ↗
DeepSeek founder Liang Wenfeng decided to break the lab's longtime "no funding, no commercialization, no roadshow" stance after watching Anthropic's Claude Mythos demonstrate capabilities trained on massive computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. and data, QbitAI reported. Anthropic released its first Mythos preview in April; DeepSeek's funding rumors emerged the same month, and the round has since closed at roughly $7.4 billion. Every DeepSeek department will at least double headcount, with the newly stood up Harness AI agentAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. team already interviewing daily for engineers to convert DeepSeek's models into autonomous agents. The pivot puts DeepSeek into open dependence on outside capital for the first time, pulling the lab squarely inside the U.S. Commerce Department export control regime that already denies it access to Nvidia's frontier GPUs.
Read at QbitAI ↗
Chinese AI models held the global lead in API call volume on OpenRouter, the AI model marketplace, for a ninth consecutive week through June 22-28, with DeepSeek-V4-Flash, MiMo-V2.5 and MiniMax M3 topping the charts, Pandaily reported. Chinese models drew 20.39 trillion tokens that week against 4.25 trillion for U.S. models, with platform-wide volume at 46.7 trillion. The U.S. share of OpenRouter token requests has slid from 72% to 33% over the period covered, citing ExponentialView research. The shift extends the picture AIPD's June 8th edition flagged at six weeks. Chinese open-weight modelsOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. such as DeepSeek, GLM and MiniMax are pulling developer demand from U.S. counterparts at a pace current Commerce Department export control design was not built to address.
Read at Pandaily ↗
Chinese AI startup Z.ai, formerly Zhipu AI, released its open-weightOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. GLM-5.2 model, with some researchers claiming parity with Anthropic's Mythos in bug finding and cybersecurity scenarios. GLM trails U.S. frontier labs on broader general purpose tasks, per The Verge. CNBC said GLM-5.2 makes open source AI a real contender on intelligence-per-dollar against Anthropic and OpenAI. The WSJ reported the assessment is fueling concern in Washington that the clampdown on top U.S. AI is handing Beijing a cyberwarfare advantage.
Read at The Verge ↗ • Read at CNBC ↗ • Read at WSJ ↗