Meta releases Muse Spark Open 816B open-weight model

Meta released Muse Spark Open, an 816B open-weight multimodal MoE it calls the top US open-weight model, with Mac and API access.

Meta released Muse Spark Open on 15 October 2026, an open-weight, natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active, and a one-million-token context, aimed at coding, agentic work, and long-horizon reasoning across text, image, video, and audio. Meta calls it the most capable US open-weight model and the first in its class to run locally via Apple MLX; Hacker News and The Decoder say that uses two M5 Ultra Mac Studios with up to 512 GB of unified memory each, while TechCrunch says local use needs at least 1 TB of RAM and works best on a single M5 Ultra Mac Studio. Weights are on Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry, with APIs at Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together AI; Latent Space reports launch pricing of $1.25 and $4.25 per million input and output tokens. License accounts differ: Hacker News cites a Muse Model License with limited commercial terms after safety post-training that met Meta’s Superintelligence Scaling Framework thresholds, while The Decoder and Latent Space describe a free or custom license requiring a “Built with Muse Spark Open” label and a separate agreement above $20 million in yearly revenue. On rankings, The Decoder reports Meta’s claim of leads over Qwen 3.8, GLM 5.3, and Kimi K3 except in coding, and a wide gap behind paid Muse Wonder Full models, whereas Latent Space cites an Artificial Analysis score of 39, first among US open-weight models but behind GLM-5.3 at 45 and Kimi K3 at 44. Hacker News also cites Helion kernels claiming up to 3.95× speedups and 3% end-to-end gains on AMD MI350.

• Meta released Muse Spark Open on 15 October 2026: an open-weight, natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active, and a 1 million-token context across text, image, video, and audio.
• Meta calls it the most capable US open-weight model and the first in its class to run locally with Apple MLX, after joint Meta-Apple work.
• Hardware accounts disagree: Hacker News and The Decoder cite two M5 Ultra Mac Studios with up to 512 GB unified memory each; TechCrunch says local use needs at least 1 TB of RAM and works best on one M5 Ultra Mac Studio.
• Weights are on Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry; APIs are at Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together AI. Latent Space lists launch pricing of $1.25/$4.25 per million…
• License descriptions differ: a Muse Model License with limited commercial terms versus a free/custom license requiring a "Built with Muse Spark Open" label and a separate agreement above $20 million yearly API revenue. Meta says safety…
• Rankings differ in framing: The Decoder reports Meta’s claimed leads over Qwen 3.8, GLM 5.3, and Kimi K3 except in coding, and a large gap versus paid Muse Wonder Full; Latent Space cites an Artificial Analysis score of 39, behind GLM-5.3…
• Hacker News cites Helion kernels claiming up to 3.95× speedups and 3% end-to-end gains on AMD MI350; TechCrunch says Meta demonstrated the model with Muse Code.