AI News Brief

🚀 Kimi K3: Moonshot AI releases world's largest open-weight model at 2.8T parameters

Moonshot AI launched Kimi K3 on July 17, a 2.8-trillion-parameter open-weight model with a 1-million-token context window and native multimodal capabilities - the largest open-weight AI model ever released. It topped the Frontend Arena coding benchmark with 1,679 Elo (1,757 votes), beating Anthropic's Claude Fable 5 (1,631) and OpenAI's GPT-5.6 Sol (1,618), and led on Program Bench and SWE Marathon. Pricing is $3/M input tokens ($0.30 cached) and $15/M output - well under frontier rates. Full weights drop July 27. Demand was so intense Moonshot paused new subscriptions.

🐉 Alibaba previews Qwen3.8 Max, a 2.4T-param open-source challenger to Fable 5

Two days after Kimi K3, Alibaba previewed Qwen3.8-Max at the World AI Conference in Shanghai - a 2.4-trillion-parameter model it ranks second only to Anthropic's Fable 5, calling it the flagship for coding and AI agents. Alibaba says it plans to release the weights as open source. Notably, Alibaba holds a 36% stake in Moonshot AI, linking the two open-weight pushes.

🛡️ Hugging Face breached by autonomous AI agent, turns to open-weight GLM 5.2 to fight back

Hugging Face disclosed that an autonomous AI agent breached production via a malicious dataset, accessing internal data and service credentials. When commercial frontier models refused to assist in defense, Hugging Face deployed Z.ai's open-weight GLM 5.2 (released mid-June, on-par with Claude Opus 4.8 and GPT-5.5) to fend off the attacker - a striking real-world case for permissionless open models.

⚖️ Trump administration weighs restricting Chinese AI models, sparking open-source fears

The Trump administration is reviving plans to restrict Chinese AI models - adding labs to the Entity List, issuing security advisories, and leveraging federal procurement rules - following Kimi K3's release. AI czar David Sacks pushed back, warning that "the leading closed labs... want the government to eliminate their open-source competition." OpenAI's Head of Strategic Futures Dean Ball separately called open-source AI a strategic risk, deepening the open-vs-closed rift inside the administration.

🧊 Google builds "Frozen v2" chip that hardwires Gemini for up to 10x inference efficiency

Google is developing a new server chip, internally dubbed "Frozen v2," that integrates Gemini's architecture directly into silicon to serve the model far more efficiently - reportedly up to a tenfold inference gain. It will run alongside, not replace, Google's TPUs, with deployment potentially beginning in 2028. Alphabet shares jumped roughly 3% on the news; the move underscores a shift toward model-specific AI silicon.

🏭 Z.AI completes giant all-Chinese-chip data center to train frontier AI

Z.AI has finished construction of a giant data center housing only Chinese-made chips - a major milestone in Beijing's push to replace restricted Nvidia silicon for AI training. The facility is meant to support development of open models like GLM 5.2 entirely on domestic hardware, narrowing dependence on U.S. export-controlled GPUs.

☁️ Microsoft expands Azure AI with AMD Helios, reportedly tests Kimi K3 for Copilot

Microsoft added AMD Helios rackscale infrastructure to Azure for large-scale AI inference, broadening compute options beyond Nvidia. Separately, Microsoft is reportedly testing China's Kimi K3 for use in Copilot and Azure - a striking signal that a leading U.S. hyperscaler is evaluating an open-weight Chinese model for production even as Washington debates restricting it.