Meta Open-Weights Muse Spark Open: 816B Parameters Run on Two Mac Studios, Tops US Open Models

Meta Superintelligence Labs released the weights of Muse Spark Open on October 15, 2026: an 816-billion-parameter mixture-of-experts model with 42 billion active parameters per token, a one-million-token context window, and native multimodal input covering text, image, video, and audio. Meta bills it as the most capable US open-weight model to date — 1,511 points on Arena's Code Arena WebDev leaderboard (highest US open model) and 89.2 on the 1,000-task MCP-Atlas tool-use benchmark, ahead of Qwen3.8 (84.5), GLM 5.3 (84.2), Kimi K3 (82.3), and Beam (78.7). The standout demo is desktop execution: the full model runs on two Apple M5 Ultra Mac Studios with 512GB unified memory each, coordinated as a single ~1TB system through Apple's MLX framework; Apple's own benchmarks the same day measured 36+ tokens/sec decode on a four-machine cluster at 8K context. Weights are on Hugging Face in BF16 and NVFP4 under the Muse Model License, with day-zero API access via Baseten, Fireworks, Groq, Modal, OpenRouter, and Together. Caveats: the license carries attribution and usage policies plus a separate agreement for MaaS businesses above $20 million revenue, and Artificial Analysis ranks it tenth among open-weight models overall — all nine ahead are Chinese.