AI News Brief

🧠 Thinking Machines Lab Releases Inkling - 975B Parameter Open-Weight Model Under Apache 2.0

Mira Murati's Thinking Machines Lab released Inkling, a 975-billion-parameter multimodal open-weight model - the largest American open-weights model to date. Released under Apache 2.0 on Hugging Face, it processes text, images, and audio, with an NVFP4 quantized version capable of running on half the GPUs. It scores 73.3% on MMMU Pro vision benchmark (vs. Claude Fable 5's 84.2%) and matches Nvidia's flagship at roughly one-third the token cost.

🚀 Moonshot AI Launches Kimi K3 - World's Largest Open-Weight Model at 2.8 Trillion Parameters

Chinese startup Moonshot AI released Kimi K3, a 2.8-trillion-parameter (50B active) open-weight model with a 1M context window - the largest open-source AI model ever released. Independent evaluator Artificial Analysis places it near Claude Fable 5 and GPT-5.6 Sol on coding and reasoning benchmarks, beating Opus 4.8 and GLM-5.2 in several tests. It debuted at #1 on LMArena's Frontend Code Arena. Open weights will be available by July 27.

⚡ Nvidia Unveils Cosmos 3 Edge and 140MW Rubin AI Factory in Japan

Nvidia announced Cosmos 3 Edge, a new AI model, and partnered 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 AI infrastructure. The announcement came as Nvidia also cut over half of its Asian AI chip buyers following expanded BIS export control compliance across the Asia-Pacific supply chain.

🔧 xAI Releases Grok 4.5 - Built for Coding, Agentic Tasks, and Knowledge Work

xAI released Grok 4.5, described as its smartest model yet, optimized for coding, agentic tasks, and knowledge work. The release intensifies competition in the frontier model space alongside OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5.

📉 Google Hits Compute Wall - Gemini Model Release Delayed, Alphabet Stock Falls 4%

Bloomberg reports Google engineers are hitting internal AI compute capacity constraints, causing delays to a Gemini model release. The setback has frustrated researchers and managers concerned about losing market position to Anthropic and OpenAI. Alphabet stock fell 4% on the news, despite the company guiding $180B-$190B in capex this year.

🛡️ Researcher Poisons Open-Weight AI Model in Under an Hour for Under $100

Katie Paxton-Fear, a cybersecurity lecturer at Manchester Metropolitan University and Semgrep security advocate, demonstrated installing a backdoor in an open-weight AI model in about an hour for under $100. The proof-of-concept highlights supply-chain security risks as open-weight model adoption accelerates - a critical concern for the open-source AI community.

🍏 Apple Hunting for AI Chip Acquisitions as Server Performance Lags

Apple is in talks with bankers and semiconductor startups to acquire AI chip companies, aiming to build server chips to compete with Nvidia. The push comes as Siri's Gemini-powered workloads currently depend on Nvidia chips in Google Cloud, exposing Apple's lack of in-house AI server silicon. Meanwhile, startup PrismML claims it shrunk a 54GB model to under 4GB (93% reduction), running 27B parameters on an iPhone 15 - and Apple wants to talk.