AI News Brief

🚀 KIMI K3: MOONSHOT AI DROPS 2.8-TRILLION-PARAMETER OPEN-WEIGHT MODEL, TOPS CODING ARENA

China's Moonshot AI released Kimi K3 on July 16-17, a 2.8-trillion-parameter sparse mixture-of-experts model (activating ~50B params) with a 1-million-token context window - the largest open-weight model ever released. On Arena.ai's Frontend Code arena, K3 scored 1,679 Elo across 1,757 votes, beating Anthropic's Claude Fable 5 (1,631) and OpenAI's GPT-5.6 Sol (1,618); it also outperformed Opus 4.8 and GLM-5.2 on Moonshot's own benchmarks. The release triggered a "Kimi moment" market selloff reminiscent of DeepSeek's January 2025 debut. Full open weights are expected July 27 under a Modified MIT license; the model is free to try now on kimi.com.

🌐 ALIBABA UNVEILS QWEN3.8 MAX PREVIEW - 2.4T PARAMETERS, GOING OPEN-WEIGHT

Alibaba launched a preview of its flagship Qwen3.8 Max model on July 19, describing it as second only to Anthropic's Fable 5, with 2.4 trillion parameters. Alibaba shares rose as much as 5.4% on the announcement, and the company expects to release open weights soon. The same day, Alibaba open-sourced its AI chip software stack and launched a "Zhenwu" cloud supernode service, signaling a full-stack open push spanning models, software, and infrastructure.

🤖 INKLING: MIRA MURATI'S THINKING MACHINES LAB SHIPS FIRST OPEN-WEIGHT MODEL

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling - a 975-billion-parameter open-weight multimodal model trained on 45 trillion tokens to understand text, image, video and audio, with a 1-million-token context. Released under an enterprise-friendly Apache 2.0 license, it is the largest US open-weight AI to date and scores 97.1% on the AIME 2026 benchmark. The lab positions Inkling as a foundation for customization rather than leaderboard dominance, betting that fine-tuned domain-specific models beat general-purpose ones.

⚡ FRONTIER AI REGULATION: DEEPMIND, OPENAI AND ANTHROPIC CONVERGE ON RULES

Over five weeks, the chiefs of Google DeepMind, OpenAI and Anthropic agreed the AI frontier needs regulating - a rare consensus among the biggest rivals. DeepMind CEO Demis Hassabis proposed a FINRA-style independent standards body to test frontier models and develop release best practices, urging it be stood up "before year end." Anthropic is separately pushing US states to adopt tougher rules, arguing 2025 transparency bills are already outdated. The White House is also weighing a dedicated AI regulator, while the Trump administration considers stricter rules on Chinese open-source AI following Kimi K3's release.

🏭 NVIDIA AND JAPAN UNVEIL WORLD'S FIRST NATIONAL AI INFRASTRUCTURE

Nvidia announced a partnership with Japan's Noetra Corp. to build a 140-megawatt "AI factory" packing 27,500 Rubin GPUs and 13,750 Vera CPUs - billed as the world's first national-scale AI infrastructure. The project signals a shift of AI compute into sovereign infrastructure, with Japan positioning the Rubin-based facility as a strategic national asset. Separately, TSMC committed to a $100 billion Arizona investment to meet "strong, multi-year" AI chip demand even as Kimi K3's launch triggered a sell-off in Nvidia and Micron shares.

🤖 GROK 4.5 TOPS VULCANBENCH CODING BENCHMARK AT 91%; PRICED AT $2/$6 PER MILLION TOKENS

SpaceXAI released Grok 4.5, its smartest model built for coding, agentic tasks, and knowledge work, priced at $2 per million input tokens and $6 per million output tokens - roughly twice the token efficiency of leading competitors. On the new VulcanBench coding benchmark, which evaluates AI agents on real-world pull-request-style tasks across Python, Rust, TypeScript, and JavaScript in reproducible Docker-sandboxed runs, Grok 4.5 topped the field with a 91% score.

📊 MOZILLA'S STATE OF OPEN SOURCE AI: OPEN MODELS CAPTURE 45% OF TRAFFIC, JUST 4% OF REVENUE

Mozilla released its inaugural "State of Open Source AI" report, finding that open-weight models have nearly matched closed models in capability and now account for 45% of traffic on routing platform OpenRouter - but capture only 4% of global AI revenue. The report concludes the real fight has shifted to the "harness" layer (the agent and deployment stack), and that deployment, support, and governance remain the main barriers keeping businesses on proprietary systems. Mozilla framed the effort as building a "Rebel Alliance" against concentrated tech power.