AI News Brief

Mira Murati's Thinking Machines releases Inkling, a 975B-parameter open-weight model under Apache 2.0

Former OpenAI CTO Mira Murati's startup released its first model, Inkling, a 975-billion-parameter open-weight multimodal model trained on 45 trillion tokens of text, image, audio, and video. Released under a permissive Apache 2.0 license, it is the largest American open-weights model to date, comparable to DeepSeek V4, GLM 5.2, and Kimi K2.6 in size. Thinking Machines also released an NVFP4 quantized version capable of running on half the GPUs and previewed Inkling-Small with 12B active parameters. The company raised a record $2 billion seed round at a $12 billion valuation in 2025.

GLM-5.2 nearly matches Claude Opus 4.8 on coding benchmarks at roughly one-sixth the cost

Z.ai's open-source GLM-5.2 scored 74.4 on FrontierSWE (independently evaluated by Proximal), within one point of Claude Opus 4.8's 75.1 and ahead of GPT-5.5's 72.6. On Terminal-Bench 2.1 it hit 81.0, up from GLM-5.1's 63.5 and within reach of Claude Opus 4.8's 85.0. GLM-5.2 captures 40% of developer tokens and costs roughly 6x less than GPT-5.5 for equivalent workloads, with API pricing dramatically undercutting Western proprietary models at $5.00/$30.00 for GPT-5.5 vs GLM-5.2's far lower rates.

US startups migrate to Chinese open-source AI as costs surge, with 41% of global downloads now from China

San Francisco-based Lindy.ai migrated 100% of its traffic from Anthropic's Claude to DeepSeek-V4, with CEO Flo Crivello saying the open-source scene is "absolutely dominated by the Chinese" and "not even close." Chinese open-source models now account for 41% of global Hugging Face downloads, surpassing US models for the first time. DoorDash, Airbnb, and Siemens are among major companies adopting Chinese AI tools, and on OpenRouter the top six most popular models all come from Chinese institutions including Tencent, Xiaomi, DeepSeek, MiniMax, and Zhipu AI.

AI startup Reflection signs $1 billion compute deal with Nebius to build a Western open-source frontier model

Brooklyn-based Reflection, founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, announced a $1 billion compute purchase from neocloud Nebius to train open-source models. CTO Antonoglou said the company plans to release its open source model later this year, aiming to "ensure there is a competitive frontier open lab in the Western world." Reflection started training models in October 2025 after initially building AI coding agents using reinforcement learning, and has positioned itself as an American answer to China's open-source dominance.

DeepMind CEO Demis Hassabis proposes FINRA-style AI standards body to regulate all frontier models

Google DeepMind CEO Demis Hassabis called for a new regulatory body modeled on FINRA (Financial Industry Regulatory Authority) to test frontier AI models and develop release best practices, funded by the AI industry, staffed by technical experts, and answerable to the US government. Writing that we may be standing in "the foothills of the singularity," Hassabis said the rules should apply to all frontier-class models "no matter their country of origin or whether they are open or closed," and urged establishment before year end. The proposal comes as 16 Nobel Laureates separately signed a Stanford campaign warning AI could drive economic transformation larger than the Industrial Revolution.

TSMC pledges additional $100 billion in Arizona, posts record Q2 profit as AI chip demand surges

TSMC announced a further $100 billion investment in Arizona on top of its previously announced $165 billion, after posting a 77% jump in Q2 profit to a record $22 billion (T$706.6 billion), beating market forecasts of T$632.6 billion and marking its ninth straight quarter of double-digit growth. The company raised its 2026 capital expenditure guidance to $60-$64 billion, up from $52-$56 billion. Separately, ASML raised its 2026 sales forecasts and pledged a capacity boost to ease fears that production bottlenecks could slow the AI boom.

Apple explores AI chip startup acquisitions as next-gen server chip Baltra slips past 2026 debut

Apple has been in talks with bankers and semiconductor startups about acquisitions to bolster its AI server capabilities, as its next-generation server chip code-named Baltra has slipped past its planned 2026 release. Apple currently relies on M2 Ultra-based systems for some AI processing while routing more demanding tasks to Nvidia GPUs in Google Cloud powering the Gemini-based Siri. The company recently acquired Israeli AI startup Q.ai for nearly $2 billion and has approached multiple semiconductor startups to gauge their interest in selling.