Meta's Muse Spark Open: 816B Parameters, Built for Macs
Meta shipped a second frontier-adjacent model in as many weeks: Muse Spark Open, an open-weight release the company bills as the "most capable" open-weight model in the United States, per TechCrunch's report. The spec sheet is where it gets interesting.
The numbers
- 816 billion parameters, open-weight — this is not a laptop model by any stretch.
- 1 million token context window.
- Natively multimodal, tuned for complex coding and agentic work.
- Designed to run locally on Apple computers — with hardware requirements to match. "Runs on Macs" means Mac Studio-class muscle, not your MacBook Air.
Read the adjective, not the headline
"Most capable open-weight in the US" is Meta's own evaluation, and every word is load-bearing: open-weight excludes closed frontier rivals, and in the US politely sidesteps the Chinese labs that currently dominate that particular leaderboard. Neither is dishonest — both are strategy.
Why the Mac angle matters
Open-weight releases usually whisper "data center." Whispering "Apple silicon" instead is a positioning shot at the local-inference crowd, and it dovetails with the local-vs-API split we laid out in our open vs. closed primer. If you run evals on it this week, the interesting question is not "is it the most capable" — it is "what does 816B open actually buy me at what hardware cost." Report back to model coverage; we will do the same.