Weights Meta had kept proprietary since April are now out: Muse Spark Open, which the company bills as the strongest open-weight model yet from a US lab.
Its design pairs 816 billion parameters under a mixture-of-experts scheme, with only about 42 billion activating per token. Text comes in alongside images, audio and video, and prompts can run to a context window of one million tokens.
Two Mac Studios, one linked system
Rather than the data center, Meta picked desktop hardware for one test, chaining a pair of Mac Studio machines with M5 Ultra chips through Apple’s MLX framework into one unit sharing a terabyte of memory. Over that rig, Muse Code tackled bug fixes and fresh features for Ice Cubes, the open-source Mastodon client, while the model stepped through footage frame by frame to choose shots for vertical video.
It traces to Muse Spark 1.2, shipped in August, before safety post-training was added so the model stands alone without Meta’s guardrails.
A split benchmark picture
A 61.8 on Meta’s JobBench agent test is more than twice Inkling’s 31.1, the mark set by Thinking Machines Lab’s US open model; Nvidia’s Nemotron 3 Ultra landed farther back at 27.2. Terminal-Bench 2.1 went to Kimi K3 at 85, and GLM 5.3 alongside Qwen3.8 cleared the 76.4 that Muse Spark Open managed, leaving Chinese entries ahead on coding.
Tool use flips it, and Meta leads: on MCP-Atlas the model picks which of 220 tools across 36 live servers to invoke, and in what sequence, scoring 89.2 to outrank every rival listed.
Access is free under the new Muse Model License, yet strings attach: products built from it must carry a “Built with Muse Spark Open” label, and firms selling it as a hosted service past $20 million in annual revenue need a separate deal.
