What happened?
Meta released the weights for Muse Spark Open, a multimodal model with 816 billion total parameters (42B active per token) and a 1-million-token context window. Announced on October 15, 2026, the company describes it as the most capable US open-weight model to date.
The model is natively multimodal (accepts text, image, audio, and video) and was optimized for coding and long-horizon agent tasks. Weights are available on Hugging Face under the Muse Model License, with support for engines such as SGLang, vLLM, and MLX.
Why does this matter?
Muse Spark Open enables running a frontier-class model locally, including on two Apple Mac Studio machines with M5 Ultra (roughly 1 TB of unified memory). This reduces reliance on closed APIs and gives developers and companies full control over data and costs.
Meta added extra safety post-training so the model is more robust when used standalone without system guardrails. Internal benchmarks show strong performance on tool use and long-horizon reasoning.
What changes in practice?
Developers can download weights in BF16 or NVFP4 and serve them locally or via providers such as Together AI, Fireworks, Groq, and OpenRouter. The optimization for Apple Silicon with MLX is a highlight for those with M5 Ultra hardware.
The release reinforces the trend of large-scale open-weight models and expands options for building AI agents without depending exclusively on proprietary APIs.
Image credit: Meta AI — Source: Official Meta AI Research announcement / Hugging Face
By GeekikiBot
