The largest set of open weights released by an American company arrived on 15 October. It is not open source, and whether you may use it commercially depends on how much money you make.

Muse Spark Open is a mixture of experts with 816B total parameters and 42B active per token. It takes text, images, video and audio in, returns text, and holds a 1M token context. The weights are on Hugging Face. The licence is Meta's own Muse Model License, which grants fine tuning and distillation rights and asks two things in return: the words Built with Muse Spark Open displayed in your interface or documentation, and, if you offer AI models as a service with more than $20m of gross revenue in any rolling twelve months, a separate written agreement with Meta. Internal use and ordinary products that are not themselves a model API are outside that clause. The BF16 checkpoint is about 1.5TB, served across four nodes of eight H200s; a 4-bit build runs on two Mac Studios.

Why this one is different

The open weights race has been Chinese for a year: DeepSeek, Qwen, Kimi, GLM, MiMo, all under MIT or Apache. This is the largest American answer to that, and it answers in a different language. MIT asks nothing. The Muse licence asks for attribution and, above a revenue line, a conversation. That is a conditional grant rather than an irrevocable one, which matters most to exactly the companies big enough to compete with Meta.

The weights are free. The licence has a revenue trigger.

How we got here

1. 21 Sep 2026Xiaomi open sources MiMo-V2.6, a trillion parameters, under the MIT licence.
2. 20 Sep 2026Qwen ships Qwen-Image-2.1 and drops Apache for a research-only licence.
3. 5 Oct 2026Reflection announces Beam, 501B, with Apache 2.0 promised but no weights yet.
4. 15 Oct 2026Muse Spark Open, 816B, under a licence with a revenue trigger in it.

What it does and does not mean

Meta's own benchmark tables are contested, and at least one comparison puts Muse Spark Open behind Kimi K3, GLM 5.3 and Qwen3.8 on Terminal-Bench 2.1, with a full evaluation report still to come. The hardware is the other limit: 1.5TB of weights is a cluster, and the two Mac Studios headline is a quantised build, not the model. What the release does change is the shape of the argument about openness. The question is no longer whether an American lab will publish weights at this scale, but what it will ask of you in exchange, and the answer here is attribution and a revenue line.