What "Open Source" Actually Includes for an AI Model

"Open source" on a model page can mean a downloadable file under any terms, a licence the Open Source Initiative lists, or a release complete enough to rebuild. Those are three different things, and a model can have any one of them without the other two. A team that depends on an open model as its fallback needs to know which of the three it actually has.

This census reads the primary for each: the licence file rather than the repository tag, the Open Source Initiative's own list rather than a vendor's adjective, and the model card and repository for what else was published. Sixteen rows covering nine publishers are in the grid. Where a publisher does not say something, the cell says "not stated", which is not a synonym for no.

Each term has a test a reader can run, and each test looks at a different artefact. Run all three before calling anything open.

The Open Source Initiative check used here is the defensible one: does the licence's identifier appear in OSI's published list, and does its page on that site resolve. Apache-2.0 and MIT pass both. OpenMDW, the two Qwen licences, the Kimi K3 licence, the LFM Open License and the Llama 4 licence fail both. The list does not distinguish a licence that was never submitted from one under review, so "not listed" is all this page claims.

Every licence cell below comes from the licence file in the weights repository or the publisher's own licence page, read in full. Two exceptions are marked: Qwen3.8-27B's Apache-2.0 comes from card metadata because the file fetch returned a redirect, and Inkling's Apache-2.0 is declared in card metadata with no licence file in the repository. Two rows have no licence file at all, and that is recorded rather than filled.

Four rows deserve a sentence each. Z.ai's GLM-5.3-Flash licence is unmodified MIT, which is as clean as the grid gets; Z.ai's own documentation calls the model "the first open-source frontier model to combine sparse and linear attention", and that is Z.ai's claim, quoted, not ours. The Qwen3.8-Max licence requires a separate licence from Qwen before any commercial use by a model-as-a-service or AI work-assistant business whose group revenue passes US$50 million in any twelve months; the Community licence on Qwen3.8-Flash-Next has the same clause with no revenue floor. NVIDIA's OpenMDW-1.1 grants permission to deal in the model materials without restriction and says it imposes no obligations on outputs, and it is still not on the OSI list. Meta's Llama 4 licence requires a licensee above 700 million monthly active users to request permission that Meta may grant at its sole discretion.

Two rows pair a permissive licence with a separate document. Muse Glimmer ships unmodified Apache-2.0 beside a usage policy with a prohibited-uses list and an under-18 exclusion; Inkling declares Apache-2.0 in card metadata with no licence file in the repository and links an acceptable-use policy. Neither publisher states how the second document relates to the licence, so this page records both files and adjudicates nothing. Muse Spark 1.3, the closed model Muse Glimmer is distilled from, has no licence file because it has no weights; what Meta's own pricing table does state is that the cheaper Contributor tier's traffic is "Used to improve our products" at $0.10 input and $0.20 output per million tokens, while the standard tier at $1.25 and $4.25 is "Not used to improve our products". The discount is paid in data.

A licence governs the weights. Whether anyone could rebuild or re-verify the model depends on what else the publisher put in the repository. Nine publishers in seven lines, from their own cards and repositories; the Nemotron 3.5 Lightning card and the gpt-oss-120b card are the two ends of the range.

A router's free variant of a model looks like an openness signal. On the day of writing OpenRouter carried 21 routes with a free suffix, and reading them against the publishers' repositories gives four flat statements.

A free route does not imply open weights: three of the 21 run models with no public weights anywhere. A free route does not imply a permissive licence: one runs LFM2.5, whose licence withdraws commercial rights above US$10 million in revenue. A permissive licence does not produce a free route: both gpt-oss models are unmodified Apache-2.0 and every gpt-oss route on the catalog is paid. And a free route can expire: three of the 21 carried deprecation dates in September, which OpenRouter documents as the deprecation date for the model endpoint. A free route is a hosting arrangement with an end date. Our price index tracks the paid routes with their check dates.

The case for an open model in a production stack is usually resilience: if the hosted vendor changes terms or disappears, the open model is the fallback. Our second-source playbook covers the architecture; these three questions decide whether a given model qualifies, and each maps to a column in the grid.

For the hardware side of the same decision, which open models fit which machines, see our self-hosting guide. If you want the licence review and the fallback design done together, that is part of what we do under AI transformation.