Open-weight AI licences compared: what five 2026 releases permit

“Open weights” stopped being one category. Across five releases published between June and August 2026, the terms attached to downloadable weights now span unmodified MIT, Apache 2.0, a bespoke licence with revenue-triggered conditions, and — for the first time on a major release — a licence that excludes four jurisdictions outright.

What is in the index, and what is not

The inclusion rule, so a reader can check it and so the next edition can reproduce it: open-weight text, video or multimodal models whose downloadable weights were published between 1 June and 5 August 2026, where the vendor publicly characterised the release as open — plus named prior releases from the same vendors where the licence changed.

Three candidates were checked and excluded, which matters more than the inclusions:

The index

Every cell checked against a primary licence source on 5 August 2026. “None” means the document was checked and the clause is absent, not that it was unexamined.

Two things in that table are worth pausing on. The first is that the two bespoke licences gate on the same number — US$20m — but measure completely different things. Kimi’s §2 counts aggregate revenue across the licensee and its affiliates over any rolling twelve months, and only bites if you also operate a Model-as-a-Service business; MiniMax’s §IV.1 counts yearly revenue with no such qualifier. A company can sit above one threshold and below the other on identical accounts.

The second is the output clause. Four of the five releases say nothing about what you do with what the model produces. MiniMax H3 does: §V.3 forbids using H3, its derivatives, or its outputs to improve any other artificial intelligence model. That is a restriction on the product of running the model, not on the weights, and it survives wherever the outputs travel. §VI.4 separately records that MiniMax claims no rights over outputs and places responsibility for them on the user — the two clauses are doing different jobs and are easy to conflate.

What moved, per vendor

A snapshot of five licences says less than the direction each vendor is travelling. This is the part no other index has, and it is the reason the comparison is worth maintaining rather than publishing once.

The MiniMax lineage is the finding here, and it does not run one way. Between M2.7 and H3 the direction reverses twice: a non-commercial licence opens into a commercial one, then the commercial one acquires a territorial carve-out, a training restriction and a governing-law clause within about two months. A vendor is not simply “closing” or “opening”; it is adjusting different levers per release, and the only way to see it is to read consecutive documents side by side.

One detail from that lineage appears in no coverage we could find: H3 is contracted through Nanonoble Pte. Ltd. (§I.15), a Singapore-form private limited company, under the law of the Hong Kong Special Administrative Region (§IX.1). M3 names neither a governing law nor an entity form. Who you are contracting with, and under which law, changed between two releases of the same product line.

A composite release has more than one licence

MiniMax H3 is not one artefact under one set of terms. Its model_index.json shows the text encoder is a Qwen3VLForConditionalGeneration model, and Alibaba publishes Qwen3-VL-32B-Instruct under Apache 2.0. So an Apache 2.0 component ships inside a release that excludes four jurisdictions.

MiniMax acknowledges this. An Additional Note in the H3 licence file, sitting outside the nine numbered sections, states that the encoder uses Qwen3-VL-32B under the Apache 2.0 License. What the repository does not appear to contain is the Apache licence text itself: the root holds a single LICENSE file, and there is no separate Apache document and no NOTICE file in the root or in the text_encoder directory. Apache 2.0 §4(a) requires that recipients of a derivative work be given “a copy of this License”. H3’s own §III.4, meanwhile, requires that distributions to third parties be accompanied by a NOTICE text file.

The practical point for anyone building on a composite model: “what licence is this model under?” is not a well-formed question once a release bundles components. The answer is a list, and the list is not always in one place.

What the error rate tells you

Most published coverage of Kimi K3 stated that it shipped under a Modified MIT licence. It did not. That was K2 — whose licence file literally opens with the words “Modified MIT License” — and the error came from assuming the successor inherited it.

Kingy made this error too. Our 18 July analysis of K3’s economics said the weights would land under a Modified MIT licence, in a piece published nine days before they landed. They did not. That article now carries a correction, and our 27 July reading of the licence as shipped is the accurate one.

Two things make this worth more than an apology. First, we checked whether Moonshot itself had promised Modified MIT — which would make this a vendor reversal rather than a reporting failure — and found no evidence for it: the Hugging Face repository carries the bespoke licence from its initial commit with no licence-replacement commit, and the project’s GitHub README states that both the code and the weights are released under the Kimi K3 License. This was extrapolation, downstream and ours included.

Second, it is a measurement problem with a cheap fix. Checking a licence claim takes about four minutes:

1. Open the licence file in the model’s own repository, not the model card, not the announcement blog, not an aggregator. On Hugging Face that is /blob/main/LICENSE.
2. Read the first line. Licence names are load-bearing. “Modified MIT License”, “MIT License” and a vendor-named community agreement are three different instruments.
3. Search the document for four words: territory, revenue, attribution, and improve. Those four catch the clauses that actually change deployment decisions, including the output-and-training restrictions that are easiest to miss.
4. Check the predecessor. If a vendor has shipped before, open the previous release’s licence beside this one. Every finding in the table above came from doing exactly that.

What this does not establish

• This is not legal advice, and it is not a compliance assessment of any company. It reports what published documents say. Corporate structure, deployment design and where your users sit can all change the answer for a specific case.
• None of these terms has been tested. No enforcement action, and to our knowledge no published dispute, exists on any licence in this index. What a clause says and what it would do are different questions.
• Licences are amended, sometimes silently. Every row here is stamped 5 August 2026. A licence file can change without a version bump or an announcement.
• Hosted API terms are a different legal relationship and are out of scope. Using MiniMax H3 through MiniMax’s API is governed by the API agreement, not the weights licence — a distinction that matters especially for H3, whose full 2K pipeline is hosted-only and not part of the weight release.
• Dates in this space rarely reconcile. MiniMax published its H3 launch article on 31 July, dated the licence 2 August, and the repository appeared on 3 August; the vendor does not reconcile the three. Accounts of when the K3 weights went public differ between 26 and 27 July. Not yet verified — we have not resolved either to a single date and have not treated any of them as settled.
• Absence of a clause is not permission. Where this index says “none”, it means the licence does not address the point — not that the point is resolved in your favour.

Primary sources

• DeepSeek V4 Flash licence and official model card
• GLM-5.2 licence
• Leanstral-1.5 official model card and licence statement
• Kimi K3 licence and predecessor Kimi K2 Modified MIT licence
• MiniMax H3 licence, H3 component index, and predecessor licences for M2.7 and M3

How this gets maintained

The index is re-checked monthly, and re-checked out of cycle whenever a vendor in it ships a new open-weight release. Monthly rather than quarterly because the M3-to-H3 delta happened inside about two months and would have been invisible on a quarterly cadence. Each edition records the date every licence was read, and changes are logged rather than silently overwritten, so that a term that moves can be seen to have moved.

Download the underlying dataset: CSV or JSON (ZIP). Both contain the comparison fields, section references, exclusion list and primary-source URLs, and are published under CC BY 4.0.

For the wider picture of how open-weight models are positioned commercially, see our State of Open-Weight AI Models report. For what these licences mean once you try to run the weights on hardware you own, see our analysis of DeepSeek V4 Flash and Kimi K3 on DGX Spark. Our evidence labels are defined in the Kingy AI research methodology.