Meta ships Muse Spark Open, 816B params, built for agentic coding
Alexandr Wang calls it the most capable US open-weight model to date, and nobody we follow has pointed it at real code yet.
Meta released Muse Spark Open and is calling it the most capable US open-weight model to date. The published specs are 816B total parameters with 42B active, a 1M token context window, and native multimodality across text, image, video and audio. The pitch is explicitly agentic, with AI at Meta saying it is built for coding and agentic work and that it excels at long-horizon tasks, complex software engineering and multi-step reasoning.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Alexandr Wang posted the same numbers and the same framing, that it is built for coding and agentic work like long-horizon tasks, hard software engineering and multi-step reasoning.
The sober read
Nathan Lambert's reaction is the one worth holding onto. He said that if the model performs roughly like the scores indicate, adoption is going to go gangbusters, and he framed it as the first US open-weight model with near frontier agentic capabilities, with GLM 5.2 and Kimi K3 as the equivalent step change from China. Note the conditional. The claim is about what happens if the published numbers survive contact with real work.
That has not happened yet. As of the release, nobody on our watchlist has run Muse Spark Open against a real codebase and posted what came out. Every agentic claim here is a vendor claim plus vendor scores, which is the normal state of a launch day and also the reason launch day tells you very little.
For working engineers, the specific thing to watch is the gap between benchmark agentic and useful agentic. The recent pattern on this site is that models land with strong published results and then diverge wildly depending on whose repo they touch. Gemini 4 Argon topped one of Theo's benchmarks and then fell apart on another, which is the kind of split a scorecard does not show you.
The open weights are the part that matters regardless. A 42B active mixture with a 1M token context is something people can actually host and point at their own agent harness, and that is how the long-horizon claim gets tested in public rather than in a table.
First open-weigh model in the US with near frontier agentic capabilities
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.