Meta released Muse Spark Open on October 15, 2026, under a document it calls the Muse Model License. The opening paragraph reads like the MIT licence, with permission “to deal in the Software without restriction.” Five numbered conditions and an attached usage policy follow, and for a Canadian business planning to build on the model, those conditions are the operative text.
What the licence grants
The grant is broad and free of charge: use, copy, modify “(including fine-tune)”, distill, merge, publish, distribute, sublicense and sell. Meta “claims no ownership of outputs generated by the Software.” For a Canadian company that wants to run a capable model on its own hardware or cloud account, that is a better starting position than many hosted services offer.
The grant is then made “subject to the following conditions,” and three of them carry real obligations.
Condition 2: the attribution line follows derivatives
Licensees “shall prominently display ‘Built with Muse Spark Open'” in the interface of each product that uses the model, or in its documentation, website or settings menu. The obligation also covers “Derivatives,” which the licence defines as any model “created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof.” Compare it with the Open Source Definition, which rules out licences that discriminate against fields of use.
The last four words matter most. A company that uses Muse Spark Open to generate training data for a smaller in-house model has, on this text, produced a Derivative, and the attribution line travels with it. The definition reaches further than the model file.
Condition 3: the revenue threshold for model services
If a licensee or its affiliates offer AI models to third parties “via API or a similar interface,” and their combined gross revenue exceeds US$20 million in any 12-month period, they “must notify Meta and enter into a separate written agreement with Meta” before offering the model in that service. The test is total revenue of the corporate group, not revenue from the AI service. Licence terms will also shape which agents can shop on your site, so read should Canadian merchants enable AI shopping agents alongside this one.
The clause carves out “internal use, research, or any commercial activity that is not Model as a Service,” and use of the model “as a component to power features of the Licensee’s own product.” A Canadian insurer running the model inside its claims tool is outside the clause. A Canadian cloud or telecom provider reselling hosted access to the model is inside it, whatever the AI line of business earns, and the terms of the separate agreement are not public.
The usage policy is part of the licence
Condition 1 incorporates the Usage Policy, attached as Exhibit A, into the licence itself. A breach of the policy is therefore a breach of the licence itself. Several prohibited uses touch ordinary Canadian business activity:
- “Operation of critical infrastructure, transportation technologies, or heavy machinery,” listed among activities that risk bodily harm.
- “The unauthorized or unlicensed practice of any profession,” including financial, legal and medical practice.
- “Representing that the use of Muse Spark Open or outputs are human generated.”
- Failing “to appropriately disclose to end users any known dangers of your AI system.”
The first item is worded broadly. A utility, mine or logistics firm considering the model for any part of an operational system should read “operation of” closely before assuming a planning or monitoring use falls outside it.
The other reading
A reasonable counsel could say this is a permissive licence with familiar guardrails. Most acceptable-use policies ban weapons and fraud, a US$20 million threshold exempts nearly every Canadian start-up, and attribution costs one line in a settings menu. On that view, the conditions are housekeeping.
We agree on the threshold and the attribution cost. We read the definition of Derivatives and the incorporation of the usage policy differently. Together they mean that an outputs-trained model built in-house carries the same conditions, including the prohibited-use list, and that a use Meta considers to be “operation of critical infrastructure” is a licence breach rather than a policy question. Those are contract terms with consequences, and they should be read as such.
Fine Print
- Instrument: Muse Model License, version “Muse Spark Open,” with the Muse Spark Open Usage Policy as Exhibit A.
- Status and effective date: Released October 15, 2026; applies on download or use of the weights.
- Who decides or enforces: Meta Platforms, Inc. for licensees outside the European region, which includes Canada.
- Who it applies to: Anyone who uses the model or a Derivative, including models trained on its outputs.
- What changes: An attribution line in each product; notice and a separate agreement for model-service groups above US$20 million in revenue; the usage policy added to internal AI rules.
- Still undecided: The terms of the separate agreement, and how Meta interprets “operation of” critical infrastructure. Both unknown.
The licence also disclaims all warranties and states that the licensee “is solely responsible for the safety, deployment, and use of its Derivatives.” That sits alongside, not instead of, whatever Canadian privacy and consumer law already requires. More analysis is in AI Policy in Canada, and our Mistral Large 4 review covers another open-weight option. This is general information, not legal advice.
Before the first production use
There is no regulatory deadline here; the decision point is the day a team moves the model from testing into a product. Before that day, legal or procurement should record the licence version, confirm whether the group’s revenue crosses the US$20 million threshold for any model-service offering, and check each planned use against Exhibit A.
FAQ
Is the Muse Model License open source?
It grants broad free rights to use, modify and distribute the model, but adds conditions: attribution, a separate agreement for large model-service providers, and a binding usage policy. It is an open-weight licence with restrictions.
Does the Muse Model License apply to models trained on Muse Spark Open outputs?
On its text, yes. Derivatives include models trained or distilled using the software or any outputs or results thereof, and the attribution and usage conditions apply to Derivatives.
When does a business need a separate agreement with Meta?
When it offers AI models to third parties through an API or similar interface and its group’s gross revenue exceeds US$20 million in any 12-month period. Internal use and product features are excluded.
