Open weights now come with fine print. Read it before the benchmark

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

Meta's Muse Spark Open, released on 15 October, is a mixture-of-experts model with about 816 billion parameters, around 42 billion active per token and a context window of up to one million tokens. Hugging Face describes it as the most capable US open-weights model to date. The BF16 checkpoint alone is about 1.5 TB.

The Muse Model License is permissive in most respects: use, modify, fine-tune, distil and sell. It adds three conditions that matter. Every product built on it must prominently display "Built with Muse Spark Open". Any licensee offering models to third parties as a service, with group revenue above $20 million in any 12-month period, must notify Meta and sign a separate written agreement first, although internal use and features inside a company's own product are exempt. And the attached usage policy bars the "unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health", as well as discrimination in the provision of credit.

Meta's smaller sibling takes a different path. The paper for Muse Glimmer 30B, dated 19 October, describes a model built to run on a single GPU and released under Apache 2.0. Meanwhile, Mistral's Large 4 weights are still listed on Hugging Face as an upcoming release, with an expected date of 31 October.

Why it matters

For a bank or insurer, "open" used to mean you could take the weights in-house and stop worrying about the vendor. That is still mostly true here: internal use is not caught by the $20 million clause. But the attribution and usage-policy terms follow the model into any customer-facing product, and the line between a tool that helps an adviser and "unlicensed practice" of a financial profession will need a lawyer's reading.

'Open' now describes a download button, not a licence.

My view is that legal review now belongs at the start of an open-model project, alongside the benchmark, not at the end. Two models from the same company, released a few days apart, carry very different terms, and that pattern is likely to spread.

What I'd watch

• Whether Mistral's Large 4 weights arrive on 31 October, and under what licence.
• Whether other labs copy the revenue-threshold model-as-a-service clause.
• How Meta interprets the financial-profession restriction for advice tools used under human supervision.