Alibaba released Qwen3.8-27B, a free open-weight model designed to run on consumer laptops, alongside the full weights of its flagship Qwen3.8 Max. The model amassed over 3 million downloads in its first weekend, and Hugging Face's State of Open Models report says Qwen has passed Google and Meta with 3 billion cumulative downloads, becoming the default fine-tuning base for developers.
Zhipu AI (Z.ai) released GLM-5.3 as an open-weight model on August 14, reporting a 50% coding improvement over GLM-5.2. On the CyberGym cybersecurity benchmark it scored 84.5, edging Anthropic's Mythos 5 (83.8) and OpenAI's GPT-5.6 Sol (83.6). On Terminal-Bench 2.1 it beat every model including Claude Fable 5 while remaining several times cheaper, though it still trails closed models on the hardest public coding tests. Open weights are promised within two weeks.
DeepSeek shipped the production build of its flagship V4 Pro (designated 0813) on August 12, ending a nearly four-month preview period. The release adds enhanced agentic coding capabilities, an open-source evaluation harness, and ranks 9th on the World of AI Benchmark. DeepSeek also confirmed price increases for V4-Flash and V4-Pro models effective mid-August.
OpenAI has dissolved its Preparedness team, the group created to challenge model release decisions on catastrophic risks, as the company progresses toward a blockbuster public offering. The move continues a pattern of safety groups changing shape or disappearing, raising fresh concerns about governance at the most valuable private AI company.
Nvidia announced partnerships with Apollo, BlackRock, Goldman Sachs, and KKR to mobilize over $500 billion in financing for AI infrastructure, with Jensen Huang emphasizing that land, power, and data-center shells are now as critical as chips. Billionaire investor Jeff Gundlach warned the move to turn AI chips into an "investable asset class" may signal a market peak.
Guardian-seen internal documents show Microsoft has approximately 2.2 million installed AI chips, far fewer than its announced data center capacity would imply. Nvidia sells almost no information on chip volumes or buyers, making it difficult for anyone to verify whether AI infrastructure buildout is matching public claims. Microsoft's stock fell 3% on the disputed estimate.
Sources tell WIRED the White House is updating its AI framework to address open-weight models, as it grapples with regulating a technology it has tried not to regulate. The EU also postponed AI Act implementation to 2027 to boost investment and compete with the US and China, while the FTC issued a proposed policy statement on AI accuracy and output steering following Executive Order 14365.