$ today_in_10_seconds
Meta shipped Muse Spark Open, an 816B-parameter open-weight model it calls the most capable in the US, and built it to run agents locally on Apple's M5 Ultra Mac Studios. Long-horizon agentic work, with a 1M-token context, no longer requires a cloud bill or someone else's terms of service. The surprise is that Meta and Apple built it together.
$ lead_story
Meta releases Muse Spark Open, an 816B open-weight model built for agents
Meta's new model is a mixture-of-experts design with 816B total parameters, of which 42B are active per token. That sparsity is what makes local inference feasible. Add a 1M-token context window and you have a model that can hold a whole repo, a long tool-call history, and a stack of retrieved docs in one session, which is exactly where agents usually fall apart. Meta is pitching it as leading agentic capabilities on the desktop. Treat that as a claim until independent evals on real multi-step tasks arrive, but the weights are out today. For founders, it means agent products that need data residency, offline operation, or predictable unit costs now have a serious open option.
// why_it_matters: The best US open model is designed for agents running on your hardware, not for chat on Meta's servers.
$ also_worth_attention
Meta and Apple team up to run Muse Spark Open across 2-4 linked Mac Studios
According to CNBC, Meta worked directly with Apple so that users can link two to four M5 Ultra Mac Studios to run the model for coding and agentic work. CNBC frames this as a thaw between longtime rivals now that John Ternus leads Apple. The practical insight is that the deployment target is a small cluster under a desk, not a single machine. Teams should start thinking about local multi-node inference as ordinary infrastructure rather than a hobbyist project.
link →
$ bigger_picture
The trend behind the headlines
The agent stack has assumed a frontier API at the center, with everything else orchestrating around it. Muse Spark Open is a bet that the center can move on-prem. Sparse MoE keeps per-token compute low, unified memory fits giant weights, and long context replaces a lot of brittle retrieval plumbing. If the evals hold up, the moats shift from who has model access to orchestration, tooling, and evaluation, which are the layers agent builders actually control. Watch for which agent frameworks ship first-class support for local multi-Mac setups. That's where the next adoption wave will start.
exo
exo is an open-source project for running large models across a cluster of everyday devices, including multiple Macs, by splitting inference across machines. It isn't confirmed as a supported runtime for Muse Spark Open. It is still the best way to understand what pooling several Mac Studios looks like in practice before you buy a second one.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime