Meta's new open model is cheap to run. The license puts Meta's name on your product.

What happened

Meta released Muse Spark Open on October 15, 2026. Downloadable weights, plus hosted access through Together AI, Fireworks AI, Groq and OpenRouter.

Together AI's launch pricing: $1.25 per million input tokens, $4.25 per million output tokens, $0.15 per million cached input tokens. That is materially below most frontier-model API pricing. Those are provider prices, not a universal Meta rate — other hosts may list differently.

The model is commercially usable. It is not unrestricted open source. Using it means accepting Meta's Muse Model License and usage policy, and the license carries two conditions that matter more than any benchmark.

Any product or service built with Muse Spark Open — or a derivative — has to display "Built with Muse Spark Open" in its interface or accompanying material.

The second condition is a revenue trigger. A model-as-a-service company whose aggregate gross revenue, including affiliates, passes $20 million in any consecutive 12-month period must get a separate agreement from Meta before offering the model or a derivative. Internal use, research, and commercial activity that is not model-as-a-service sit outside that requirement.

What this is NOT

This is not a story about your editing workflow. If you run a Muse-powered transcription step on your own podcast and never sell access to it, nothing in the license changes your Tuesday.

It is not a ban. Most publishers can use this model commercially without paying Meta a cent.

It is not open source in the sense the phrase usually carries. Attribution is mandatory, derivatives inherit the obligations, and a revenue ceiling exists. Fine-tuning does not wash the terms off.

And it is not a settled definition. The reporting does not establish which products Meta would classify as "model-as-a-service," whether Meta offers a standard agreement for companies that cross $20 million, or what the penalty is for omitting the label. Independent quality, latency and context-window comparisons weren't established either. Treat the license text as the document, not the coverage.

Why it matters if you publish

The line falls between using a model and selling a model.

A newsroom running Muse Spark Open internally for research, headline testing or editing is probably on the safe side of the $20 million trigger — though the attribution requirement may still attach to anything reader-facing.

A creator-tool startup selling an AI writing assistant is the one that should read the license twice. That is the business the threshold was written for.

The practical consequence is design, not law. If you ship a paid AI feature — a chat box in your app, a summarizer in your newsletter product, a white-label tool you sell to other creators — Meta's brand now appears next to yours. For a white-label deal, that's a sales conversation you didn't have before. Clients who pay for "your" tool tend to have opinions about a competitor's logo on it.

The price is the reason anyone will accept those terms. Meta has been pushing its Muse line hard this month, including the agentic Muse Wonder behind a premium tier, per CNBC. Spark Open is the cheap end, and cheap compounds when you run a model across thousands of documents.

What to do about it

1. Ask your developer one question this week: does anything we sell call Muse Spark Open? Get a yes or no in writing.
2. If yes, put "Built with Muse Spark Open" in the product UI or docs now. It is the cheapest compliance item on the list.
3. Check your contracts for white-label or reseller clauses that promise unbranded delivery. Attribution may conflict with them.
4. Run the $20 million math on gross revenue including affiliates — not profit, not the AI product line alone.
5. If you're near the threshold, email Meta and ask for the separate agreement before you ship, not after.
6. Price the alternative. Log your current per-million-token spend, compare against $1.25 in and $4.25 out, and decide whether the branding condition is worth the delta.
7. Don't assume fine-tuning frees you. The license covers derivatives explicitly.