The takeaway: America finally has a frontier-scale open-weights model of its own. Thinking Machines Lab — the startup founded by former OpenAI chief technology officer Mira Murati — released Inkling yesterday (Wednesday 15 July): a 975-billion-parameter Mixture-of-Experts model with the full weights downloadable from Hugging Face under an Apache 2.0 licence. For two years the open-weights frontier has been a Chinese story — DeepSeek, Z.ai's GLM, Moonshot's Kimi. As of yesterday, there's a serious American entry — and an early argument about whether it's good enough.

America gets its open-weights giant

The specs, from the announcement: "a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio and video." Multimodal on the input side — text, images and audio in, text out — which matters below. A smaller sibling, Inkling-Small (276B parameters, 12B active), is so far only "a preview": it isn't on Hugging Face yet, so the open-weights claim belongs to the big model alone.

The licence is the real story, and we checked it at the source. The Hugging Face repository is tagged Apache 2.0 — a genuinely permissive open-source licence, not the restricted "open-ish" licences of the Llama era. One rider: the weights come with a separate acceptable-use policy asserting that "By accessing, downloading, or using any Model Materials, you agree to be bound by this Model AUP" — prohibiting surveillance, automated rights-affecting decisions and safety-measure circumvention. Purists will argue about how that click-through interacts with Apache 2.0. It is still one of the most open releases ever made at this scale by a US lab.

The benchmarks are the company's own until independents finish their runs, so label them accordingly: Thinking Machines reports 97.1% on AIME 2026, 87.2% on GPQA Diamond, 46% on Humanity's Last Exam with tools, and 77.6% on SWE-Bench Verified — a table that, unusually, includes rows where its own model trails competitors. The first independent read arrived within hours: Artificial Analysis scores Inkling 41 on its Intelligence Index, tenth of the 97 comparable models it tracks, and calls it "amongst the leading models in intelligence, but particularly expensive when comparing to other open weight models of similar size" at launch pricing of $1.87 in / $4.68 out per million tokens.

That cost point is exactly where the Hacker News debate (750+ points) landed. "Its not as good as GLM 5.2 for agentic workflows while also being bigger," ran the most-discussed early verdict. The counter-argument: "GLM 5.2 underwent extensive post-training and iteration since its original release to reach its current state. This seems like an extremely strong model for a first release, with a lot of potential for improvement." And a third camp noted the honesty: "Then why are they publishing the benchmarks which makes them look worse than GLM 5.2?" And nobody in the thread disputed the differentiator — as one commenter put it, the "largest open weight model that supports audio", in a field where the Chinese giants are text-first. TechCrunch frames the release the way the company itself does: not a claim to the strongest model, but a bet that businesses want capable open weights they can tune — via its Tinker fine-tuning platform — rather than rent a sealed one-size-fits-all frontier model.

xAI opens Grok Build's source — mid-purge

Quiet follow-on to the story we've tracked all week: xAI has open-sourced Grok Build. The repository went up on Tuesday 14 July under — again — Apache 2.0, and had climbed past 4,800 GitHub stars by this morning as it hit Hacker News's front page overnight. xAI hasn't said the release is a response to the wire-level teardown that caught the CLI uploading entire repositories, or to Elon Musk's purge promise that followed — but the timing speaks: the upload path that researchers had to reverse-engineer from network traffic is now inspectable source code. What's still missing is what we noted yesterday: a formal incident report, a deletion timeline, and any way for affected users to verify the promised purge. We'll be reading the source.

Xi takes the WAIC stage tomorrow

Tomorrow (Friday 17 July) Xi Jinping delivers the opening keynote of the World AI Conference in Shanghai — the first time China's leader has appeared at the country's flagship AI event in person, a slot previously filled by Premier Li Qiang. The conference runs to Monday 20 July alongside a High-Level Meeting on Global AI Governance, and the expectation — as we set out on Monday — is that Xi uses it to give shape to the proposed World AI Cooperation Organisation, to be headquartered in Shanghai. We'll cover the keynote from the transcript tomorrow.

Three days of included Fable left

We re-checked Anthropic's promotional-access page this morning: the extension still ends "July 19, 2026 at 11:59:59 PM PT" — that's 07:59:59 BST on Monday 20 July. The 50% Claude Code usage-limit boost ends at the same moment. Twice now the deadline has moved in the final hours; we'll be watching Sunday evening. If you're deciding what to do when the meter starts, the guide is current.

And finally: a keyboard for your agents

OpenAI shipped its first hardware — and it's a macropad. The Codex Micro, built with keyboard maker Work Louder and priced at $230, is a 13-key desk pad whose six "agent keys" glow with the live status of your Codex threads — white idle, blue thinking, green done, red error — plus a push-to-talk key for driving agents by voice. Orders opened yesterday, limited run. The clearest sign yet of where the labs think coding is going: you don't type the code, you supervise the lights.

Also on the wire: DeepSeek's legacy model endpoints retire a week on Friday (24 July) with a V4 formal release still expected before the cut-off; the Gemini 3.5 Pro rumour mill points at this week; and in Apple v. Liu the promised preliminary-injunction motion had not appeared on the docket as of Tuesday's entries. On the OpenAI Codex accountability threads we've been tracking, there is still no visibly staff-badged reply on any of the three open issues.