The takeaway: This week put every meaning of "open AI model" on display at once: Inkling shipped genuinely Apache-2.0 weights with a click-through use policy on top, xAI published Grok Build's source with the crucial code stubbed out, and the Chinese giants it all gets compared to sit under plain MIT licences. "Open" is doing a lot of different jobs in those sentences. Here's the working guide to what the words actually mean, what each level really buys you — and the questions to ask of any "open" release.

Open weights ≠ open source

Open weights means exactly one thing: you can download the model's parameters — the billions of numbers that constitute the trained model — and run them on hardware you control. It says nothing about how the model was made.

Open source, applied to AI, is a stricter claim, and there is now a formal definition: the Open Source Initiative's Open Source AI Definition (version 1.0), which requires the freedoms to "Use", "Study", "Modify" and "Share" the system — and, critically, access to the "preferred form" for modification. That includes the training code and what OSI calls Data Information: "[s]ufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system."

By that standard, almost nothing at the frontier qualifies. Weights tell you what the model is; they don't tell you what it was trained on, or let you rebuild it. Every headline release this week — Inkling included — is open-weights, not open-source AI. That's not a scandal; it's a distinction worth keeping sharp, because the two get traded on interchangeably in marketing.

The licence spectrum, illustrated by this week

MIT — the most permissive mainstream licence. Z.ai's GLM 5.2 and DeepSeek's V4 Pro both carry a plain license: mit tag on Hugging Face: use, modify, redistribute, commercially, no strings beyond keeping the copyright notice. This is one reason the Chinese open-weights wave spread so fast through the developer ecosystem.

Apache 2.0 — equally permissive in practice, with an explicit patent grant. This week's twist: both Inkling's weights and the Grok Build repository are Apache 2.0 — but Thinking Machines pairs its weights 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, unless we have otherwise agreed in writing." The AUP prohibits things like surveillance and automated rights-affecting decisions. How a click-through rider interacts with Apache 2.0's grant is genuinely contested territory — purists say it makes the release not-quite-unconditional; in practice, it is still among the most open licences ever attached to a model this size.

Custom "community" licences — the Llama era's contribution. Meta's Llama licence looks permissive until you reach its Additional Commercial Terms: if, as of Llama 3.1's release date, your products exceeded "700 million monthly active users in the preceding calendar month", you must request a licence from Meta, "which Meta may grant to you in its sole discretion". A licence with a competitor carve-out is not open source under any formal definition — whatever Meta's marketing says — which is precisely why the tag on those repos reads license: llama3.1, not MIT or Apache.

What openness actually buys you — and its limits

The real benefits are concrete: self-hosting (your data never leaves your infrastructure — the exact concern the Grok Build teardown made vivid); fine-tuning (turn a general model into your model); permanence (no deprecation, no rug-pull, no meter starting on a promotional cliff); and auditability — up to a point.

That last one deserves this week's asterisk. Weights are not legible: you can test an open-weights model exhaustively, but you cannot read billions of parameters and see what it learned. And "source available" can be narrower than it sounds — Grok Build's repository shipped with the upload code stubbed out, a single squashed commit, and no way to file an issue. Openness is a spectrum on the transparency axis too: weights without data, source without history, repositories without a forum. Each is more open than nothing, and each can still leave the question you actually care about unanswerable.

How to read any "open" release

Five checks, one minute: (1) the licence field on the actual repository — not the press release; (2) any rider documents (AUPs, community terms) and what you agree to by downloading; (3) what's actually downloadable today — announced siblings can be "previews" that never hit the hub; (4) whether code releases carry real history and accept scrutiny, or are periodic squashed exports; (5) what still lives server-side, because no licence on the client tells you what the vendor's servers do.

The stakes are rising: the open-weights frontier is now a live US-China contest, and governments writing AI rules increasingly hinge obligations on whether a model is "open". When the word starts carrying regulatory weight, precision about what it means stops being pedantry.