AI's Most-Watched Debut Ships as Open Weights
Thinking Machines Lab — the startup Mira Murati founded after leaving OpenAI's CTO seat — released its first model yesterday, and the headline is not the benchmark table but the licence. Inkling is a natively multimodal mixture-of-experts system — 975B total parameters, ~41B active, trained on 45 trillion tokens of text, image, audio and video — and the weights are on Hugging Face under Apache 2.0, with no field-of-use restrictions, revenue caps or acceptable-use tethering. The lab's own framing is unusually honest: Inkling is "not the strongest overall model available today, open or closed" — its pitch is efficiency, with company-reported figures of 77.6% on SWE-bench Verified and 97.1% on AIME 2026, and roughly a third of the tokens Nvidia's Nemotron 3 Ultra needs to hit the same Terminal Bench score.
Two things make this more than a model launch. First, the supply side: the fresh open-weight shelf has been overwhelmingly Chinese this year — 41% of Hugging Face downloads, the top six OpenRouter models — and Inkling is the first US frontier-lab debut to enter on that side of the ledger, betting the company on the "done renting" thesis rather than arriving at it late. Second, the economics: if agentic workloads are serial, tokens-per-task is the number that shows up on invoices, and a lab choosing to compete on efficiency under a licence that lets you run the weights on your own infrastructure, in your own jurisdiction, is competing on exactly the terms European buyers have been pushed toward. The caveat stands until the independent numbers land: every figure above is vendor-reported.
Japan Buys the Full Stack
Jensen Huang spent today in Tokyo declaring "the dawn of AI in Japan" and unveiling a slate of partnerships across manufacturing, robotics, automotive and healthcare — Kawasaki, Toyota, Honda and NEC among them. The centrepiece is Project Noetra: a SoftBank-led consortium of 44 Japanese companies backed by roughly ¥1 trillion (~$6.2B) in government subsidies to build a national-level "physical AI" model — Japan's sovereign-AI strategy made concrete, with Nvidia supplying the silicon. Nikkei's framing is the honest one: Nvidia is cashing in on sovereign AI, and Japan is the latest tie-up. It follows Seoul's late-June "triple axis" plan of semiconductors, physical AI and datacenters, totalling ~$880B in mostly corporate investment.
Sovereign AI has crossed from policy paper to industrial policy in Asia, with committed capital at scales Europe has not matched: the EU's Cloud and AI Development Act aims to triple datacenter capacity over five to seven years, while a new survey warns the continent's sovereign ambitions could stall on plain infrastructure shortfalls. Worth noticing, though, what "sovereign" means in each build: in the Asian pattern it means domestic models trained by national champions on locally installed — but American — silicon, with the vendor the same everywhere. Jurisdiction over the datacenter, the training data and the operating entity is the substance; the badge on the GPU is not. That is a more attainable definition than the maximalist one, and it is the one Europe's builders should be measured against too.
Nobody Grades Above a C+
The Future of Life Institute's Summer 2026 AI Safety Index, published last week, graded nine frontier labs across 37 indicators in six domains — and the top mark in the industry is a C+, held by Anthropic. OpenAI and Google DeepMind earn Cs, Meta a D+, and xAI, DeepSeek and Mistral fail outright. The report's sharper finding is directional: the largest AI companies have weakened key safety commitments even as their models grow more capable — the voluntary-pledge system is eroding before governments have a durable replacement in place.
Read alongside the UN scientific panel's preliminary report from earlier this week, a pattern is forming: the independent assessment layer — indices, panels, third-party benchmarks — is converging on the same message from different directions, and it is becoming the de facto trust infrastructure while regulation phases in. For European enterprises there are two uncomfortable details. The 2 August GPAI obligations lean heavily on provider self-reporting, which is precisely the mechanism the FLI report says is weakening. And a European lab failing outright complicates the comfortable equation of European-equals-trustworthy: where a model is served answers a jurisdiction question, not a governance one, and buyers now need both answered — by someone other than the vendor.
Quick Hits
- Fable 5's included window closes Sunday. After three extensions, Anthropic's plan-included access to Claude Fable 5 ends 19 July at 23:59 PT, with metered usage credits at $10/$50 per Mtok from 20 July. Whether it expires on schedule or extends a third time is now a live pricing-power signal.
- The EU's AI Omnibus is still not in the Official Journal. Publication is expected mid-to-late July with entry into force three days later; meanwhile the Commission's targeted consultation on high-risk classification guidelines closed 23 June, with the guidelines themselves still pending. The 2 August GPAI penalty clock is unmoved.
- Europe's open-model July is now a two-week test. Mistral's open-weight MoE remains in gated early access with no public checkpoint, and OpenEuroLLM's first models are due 31 July amid flagged compute constraints — with Inkling now setting the bar for what a debut open-weight release looks like.