Access comes with a boundary
The terms split access to the model from the business of serving it. Anyone can use the weights for free, but API providers above the revenue threshold need Meta’s approval through a separate agreement. The required attribution also means the model’s free availability is not the same as an unrestricted offering.
The hardware claim is specific: Meta says Muse Spark Open is optimized for Apple M5 Ultra and runs locally on two Mac Studio systems with 512 GB of unified memory apiece. That signals a substantial local setup, not a claim that the model will run on an ordinary workstation.
The missing benchmark
Meta’s announcement establishes where the weights can be obtained, what hardware the company says they target, and how providers may offer the model. It does not establish how Muse Spark Open compares with Qwen, GLM or Kimi on coding or agent tasks.
I think that is the more consequential gap than the access terms. The release frames the model around programming and agents, but without comparative results, the hardware specification says more about where it can run than what it can do. In an open-weights race, availability gets a model into consideration; evidence of performance is what gives developers a reason to stay.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X