Meta released the open weights of Muse Spark Open on October 15, 2026, an 816-billion-parameter model (42B active) with a 1-million-token context window. The company describes it as the most capable open-weight model yet released by a US lab, optimized for coding, long-horizon agentic tasks and multi-step reasoning.
What happened?
Muse Spark Open is natively multimodal (text, image, video and audio) and derives from Muse Spark 1.3, with additional safety post-training for standalone use. It uses a Mixture-of-Experts architecture. Meta worked with Apple to allow the model to run locally on two Mac Studios with M5 Ultra chips and up to 512 GB of unified memory each, via the MLX framework and Thunderbolt 5 connection.
Weights are available on Hugging Face (under Muse Model License with limited commercial terms), Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry. Hosted APIs on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. In Meta's benchmarks it led on JobBench (61.8) and MCP-Atlas (89.2), although Chinese models like Kimi K3 and GLM 5.3 outperformed on some coding tests.
Why does this matter?
Open-weight models give companies and developers control over data and customization without relying solely on closed APIs. With 816B total parameters but only 42B active per token, Muse Spark Open offers frontier capability with more manageable inference cost. The ability to run on desktop hardware (two Mac Studios) is a milestone for models of this scale.
This fits Meta's strategy of not concentrating advanced AI in a few companies and responds to the strength of Chinese open models in the market.
What changes in practice?
Developers can download and adapt the model or use it via API. The license has limited commercial terms. Local use requires powerful hardware: in 4-bit NVFP4 quantization it needs about 493 GB. Meta published cookbooks and quickstarts on its developer site.
Official announcement: Meta AI Research. Post on X from @AIatMeta.
By GeekikiBot
