Open-weight licences and what you can ship

Open weights is not one licence

The phrase covers at least four different legal positions, and the difference between them is the difference between shipping and asking a lawyer. Apache 2.0 and MIT impose essentially nothing on a commercial user. A community licence permits commercial use but attaches conditions -- attribution, naming, acceptable use, sometimes a size threshold above which the permission stops applying to you specifically.

Because every model we serve has to carry an open licence to be listed at all, the catalogue doubles as a survey of what the current open-weight landscape actually permits. Here is the whole distribution.

Where the Llama licence differs from Apache 2.0

The Llama 4 Community Licence grants a broad, royalty-free right to use, reproduce and modify the model, including commercially. Three conditions come with it that an Apache 2.0 licence does not impose.

Attribution: products built on it must display Built with Llama, and the licence text and a copyright notice must travel with any redistributed copy. Naming: a model you train using Llama outputs must carry Llama at the start of its name. Scale: the grant is conditioned on your monthly active users, counted in the calendar month preceding the model's release date, staying under the threshold in the licence -- above it, you must request a licence from Meta rather than relying on the community grant.

For almost every company that threshold is theoretical. It matters because procurement review is not about likelihood, it is about whether a condition exists that someone has to sign off on -- and Apache 2.0 has none of these, which is why an Apache-licensed model of equivalent quality is often the cheaper choice in review time even when it is the same price in tokens.

The conditions worth reading before a launch, not after

Four patterns account for nearly every condition in the bespoke community licences. None of them prevent commercial use; all of them are things a launch checklist should have already resolved.

The licence is a field, not a footnote

Every model record carries its licence as structured data, so the models endpoint returns it and you can gate a deployment on it without a human reading a table. That is the difference between a compliance page and a compliance control.

The same data drives our published licence table, which lists all 24 models against their publisher and licence in one place. A procurement reviewer asking which of these are Apache 2.0 should be able to answer it in one query rather than one meeting.

# Every model whose licence is Apache 2.0 or MIT, no key required.
curl -s https://router.xark.io/api/v1/models \
  | jq -r '.data[] | select(.license == "Apache 2.0" or .license == "MIT")
           | "\(.id)\t\(.license)"'

# Fail a build if a model outside the allowlist has crept into config.
curl -s https://router.xark.io/api/v1/models/z-ai/glm-5.2 | jq -e '.license == "MIT"'

The models we will not resell, and why

A cheap-inference reseller that lists Sora or Veo is either misrepresenting what it serves or serving it in breach of someone's terms. There is no third option, and a buyer who has researched this market has already seen the warnings. So the refusal list is published with reasons attached, because naming the specific models is more convincing than a generic paragraph about compliance.

An open lineage does not make the next version open

The Wan entry on that list is the one that catches people, and it is worth stating on its own. Wan 2.1 and Wan 2.2 are Apache 2.0 -- genuinely permissive, genuinely self-hostable, and we serve 2.2 for exactly that reason. Wan 2.5 and 2.6 are API-only products from the same publisher. The weights were never released.

A developer who has been running Wan 2.2 locally under Apache 2.0 will reasonably assume the next version continues that arrangement, and a provider advertising Wan 2.5 access is not selling them an open-weight model. Check the release the version number actually refers to, not the family. Version continuity is a marketing artefact; the licence is the thing that binds.

Hosting licence and output licence are different questions

Two rights are in play whenever you call a hosted open-weight model. One is our right to serve it -- that is what the model's licence permits, and it is the reason a closed-weight model cannot appear on this platform at any price. The other is your right to use what comes out.

Our terms do not override the model's licence, and they are not written to. The model's licence governs your use of its output; where it carries an acceptable-use policy, that policy applies to you as well as to us. That is a narrower promise than a hosting provider claiming to grant you rights it does not hold, and it is the only version of the promise that survives contact with the actual licence text.

A checklist before you ship on open weights

Name the exact model version, not the family. Wan 2.2 and Wan 2.6 are different legal objects.

Read the acceptable-use policy if the licence incorporates one by reference. It is part of the licence, and it is where the prohibited-use list lives.

Record any user-count threshold next to your current number, in the same document, and set a review trigger rather than a calendar reminder.

Ship the attribution string. Built with Llama in a footer is a five-minute task that stops being a five-minute task once the product is public.

Gate it in CI. The licence is available as a field on the models endpoint, so an allowlist check is a build step rather than a memo.