On 15 October 2026, Meta released Muse Spark Open, which it called the most capable open-weight model in the United States, and said it was built to run on Macs. The footnote is the product decision: local use wants about 1 terabyte of RAM. Open-weight is not the same thing as cheap to run.
What shipped
TechCrunch reported an 816-billion-parameter model with a 1 million token context window, natively multimodal, aimed at coding and agent work. Meta’s evaluations claim it beats other popular open-weight models. Engadget’s figure, cited by TechCrunch, is at least 1 TB of memory, with the practical machine being a Mac Studio on an M5 Ultra. Meta showed it beside Muse Code on a Mastodon client — bug fixes, features, and a full build — and on cropping footage in DaVinci Resolve. Source: https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/
Three ways to use a model like this
- Local, if you have the machine. A studio workstation for a research seat, not a laptop fleet for a delivery team.
- API, if you have the volume. You pay per token and you do not own the weights. You also do not buy a terabyte of RAM per developer.
- A smaller model, if the task is narrow. Classifying a ticket or filling a CRM field does not need 816 billion parameters.
Weights you can download still need an ops plan: version, eval, and a rule for what the agent may change in the repo. That is the same bar as any other coding agent. For the assistant product these models feed, see what Meta Muse is.
FAQ
Can we run Muse Spark Open in an Ahmedabad office?
Only on a machine with memory in that class, plus power and cooling to match. For a product used by customers, an API or a smaller self-hosted model is the usual path. Do not plan a team rollout around one Mac Studio.
Does open-weight mean our data stays on our disk?
Only if inference actually runs on hardware you control. Downloading the weights and then calling someone else’s hosted copy is an API with extra steps.
Match the model to the machine
Contact Altron Technologies with your data-residency rule and your task. We will say whether local weights are worth the hardware.