Which open-weight image and video models can you use commercially? (2026)

The short answer

We read the licence file, model card and README for each model below. The table further down has the detail; here's the summary.

• Commercial use allowed, permissive licence: Z-Image and Z-Image Turbo, Qwen-Image-2512, Qwen-Image-Edit-2511, FLUX.2 [klein] 4B, FLUX.1 [schnell], Wan 2.2 (Apache 2.0); HiDream-O1-Image and LongCat-Video (MIT).
• Commercial use allowed, with use restrictions you pass on: SDXL (CreativeML Open RAIL++-M).
• Commercial use allowed below a threshold or inside a territory: LTX-2.3 and LTX-2.5 (free under $10M annual revenue), HunyuanVideo 1.5 (not licensed in the EU, UK or South Korea; licence request above 100M monthly users), MiniMax H3 (not licensed in the EU, UK, South Korea or US; authorisation above $20M yearly revenue).
• Non-commercial only: FLUX.2 [klein] 9B (FLUX Non-Commercial License v2.1) and Qwen-Image-2.1 (Qwen Research License).

The licence families, in plain terms

Permissive: Apache 2.0 and MIT

Both allow commercial use, modification and redistribution with no field-of-use limits. Apache 2.0 asks you to ship the licence text, keep notices, and mark files you changed, and it includes a patent grant from contributors. MIT asks only that the copyright and permission notice stays in copies. Neither says anything about what you generate.

Two caveats. A model card can add terms the licence file doesn't have: the Wan 2.2 README asks that use not involve sharing unlawful or harmful content, and the FLUX.2 [klein] 4B card says acceptable use for commercial open-weight users is governed by BFL's Developer Terms of Service, Self-Hosted Commercial License Terms and Usage Policy, without explaining how that sits alongside the Apache grant. And Apache never licenses trademarks, so you can't imply the developer endorses your product.

Responsible AI licences (OpenRAIL)

SDXL's CreativeML Open RAIL++-M has no non-commercial clause, but it attaches a list of banned uses (Attachment A) that travels with the model. It also counts offering the model through an API or web app as distribution, so a host must write those restrictions into its user terms as an enforceable provision. Stability's model card separately says the model "is intended for research purposes only", which the licence text itself doesn't say.

Non-commercial and research licences

These allow research, evaluation and personal use, and send everyone else to the developer for a paid licence. The FLUX Non-Commercial License v2.1 defines non-commercial as use where you receive no direct or indirect payment, and excludes revenue-generating activity and end-user-facing use. The Qwen Research License grants rights "FOR NON-COMMERCIAL PURPOSES ONLY", defined as research or evaluation.

Outputs are where these differ. The FLUX licence says you may use Output for any purpose, including commercial, except to train a competing model, though another clause bars commercial or production use of data the model produces, and the licence doesn't reconcile the two. The Qwen Research License has no clause that frees outputs for commercial use.

Community licences with thresholds

Free for most users, paid for the largest. Lightricks' LTX licences require a paid agreement for entities with at least $10M in annual revenue, counted across affiliates. MiniMax H3 needs prior written authorisation once a commercial product passes $20M in yearly revenue. Tencent's HunyuanVideo 1.5 licence uses users instead: if your products had more than 100 million monthly active users in the month before the release date, you must request a licence.

Territory exclusions

Some licences simply don't apply in certain places. The Tencent Hunyuan Community License defines its territory as worldwide excluding the EU, the UK and South Korea, and use outside it is unlicensed. The MiniMax H3 Community License excludes the EU, the UK, South Korea and the United States. If you or your users are in an excluded territory, there's no licence to comply with: you'd need a separate agreement.

Acceptable-use policies and output rules

Almost every non-permissive licence carries a banned-use list (illegal content, harming minors, non-consensual imagery, automated decisions affecting legal rights, military use in some), and most require you to pass it on to your users. Several also restrict what you do with outputs: HunyuanVideo 1.5 and MiniMax H3 forbid using outputs to improve other AI models, the FLUX non-commercial licence forbids training competing models on them, and the HunyuanVideo policy requires machine-generated content placed in public to be labelled as such.

Licence and commercial use, model by model

Notes on the tricky ones

FLUX.2 [klein]: 4B and 9B differ

The two sizes share a name and an architecture family but not a licence. BFL's README lists [klein] 4B and 4B Base under Apache 2.0, and [klein] 9B, 9B KV, 9B Base and FLUX.2 [dev] under the FLUX Non-Commercial License. If you need commercial use without a BFL contract, it's the 4B. The 9B licence also requires you to run content filters or review outputs before you distribute them.

Qwen-Image-2.1 broke the Apache pattern

Qwen-Image-2512 and Qwen-Image-Edit-2511 are Apache 2.0. Qwen-Image-2.1's card tags its licence as qwen-research, and the LICENSE file is the Qwen Research License Agreement dated September 20, 2026, limited to research or evaluation. Commercial use needs a separate licence from the Qwen team, requested by email per section 2(b). Don't assume a newer Qwen-Image release carries the older licence.

HunyuanVideo 1.5 and MiniMax H3: check where you are

Both licences grant nothing in their excluded territories, and both cover outputs too: HunyuanVideo's section 5(c) makes using or displaying the works or their Output outside the territory unlicensed. MiniMax's licence Q&A says the hosted MiniMax API is available globally and that companies in the restricted regions can apply for a formal licence, describing the limits as reflecting "today's regulatory reality". Only the 768p Base checkpoints of H3 are open; the 2K upscaler isn't released.

HunyuanVideo 1.5 also asks for things most teams overlook: give third parties a copy of the agreement, put the use restrictions into your terms of use, disclose the provider's full legal name to end users, and state that Tencent doesn't endorse the service.

LTX-2.3 and LTX-2.5 use different licence files

The LTX-2.3 card links LICENSE-2 (LTX-2 Community License, January 2026). The LTX-2.5 card links LICENSE-2_x (LTX-2.x Community License, August 2026). Both put the paid threshold at $10M annual revenue. The 2.x licence adds a transfer rule: a commercial entity receiving your fine-tune must get its own paid licence before using it, and you must tell them so in writing.

SDXL's card and licence disagree

The OpenRAIL++-M text allows hosting and commercial use with restrictions. The model card's Direct Use section says research purposes only. The licence says it "is informed by the model card", so if SDXL matters to your business, read both.

Models with no weights to license: Wan 2.7 and Wan 3.0

Wan 2.7 and Wan 3.0 are API-only on Alibaba Cloud, so there's no weights licence to read: commercial use is governed by Alibaba Cloud's service terms. The Wan models you can download, including Wan2.2-Animate and Wan2.2-S2V, are Apache 2.0 like the rest of Wan 2.2.

Open weights vs open source

"Open weights" means you can download the trained parameters. It says nothing about the terms. "Open source" has a specific meaning: the Open Source Definition forbids discrimination against persons or groups and against fields of endeavour. A licence that bars commercial use, excludes countries or charges above a revenue line fails those tests, so FLUX.2 [klein] 9B, Qwen-Image-2.1, HunyuanVideo 1.5, LTX-2.x and MiniMax H3 are open-weight but not open source by that definition. SDXL is open-weight too: its OpenRAIL licence attaches a list of banned uses, which is a field-of-use restriction.

Apache 2.0 and MIT weights pass the licence test. The OSI's Open Source AI Definition goes further and asks for training code and detailed information about the training data as well as the parameters, which almost no image or video model release includes. So "is Wan 2.2 open source?" has two honest answers: the weights are under an open-source licence; the model as a whole doesn't meet the OSI's AI definition. Our commercial use page uses the licence test and marks every catalogued model as one or the other.

For product decisions the licence matters more than the label. A permissively licensed model lets you ship, fine-tune and redistribute without a contract. Everything else needs at least a read of the conditions, and sometimes a conversation with the developer.

A checklist before you ship

1. Find the licence file for the exact checkpoint, not the model family. Base, distilled, turbo and edit variants can differ.
2. Read the model card as well as the licence. Cards add out-of-scope lists and policy links.
3. Check territory clauses against where your company, servers and users are.
4. Check thresholds: revenue (LTX, MiniMax H3) or monthly users (HunyuanVideo 1.5), counted across affiliates where the licence says so.
5. If you host the model for others, check what you must pass on: licence copies, use restrictions in your terms, provider disclosure, attribution in the UI.
6. If you plan to train on outputs, check output clauses. Several licences forbid improving other models with them.
7. If you fine-tune, check whether derivatives inherit the licence and whether transferring them triggers anything.

On Windpaint

The commercial use page lists the licence and conditions for every model in the Windpaint catalogue, each linked to its source. Today Z-Image Turbo (Apache 2.0) and FLUX.2 [klein] 4B (Apache 2.0) are live on the API; the rest of the catalogue is coming soon.