Meta releases 816B-parameter Muse Spark Open
Meta has released its largest open-weight model, pairing a 1M-token context window with coding, multimodal and long-running agent capabilities.
Meta has released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with roughly 42 billion parameters active per token. The company describes it as its most capable US open-weight model to date, targeting coding, multimodal reasoning and long-running agent work.
The release is unusually broad. Meta says the model can handle text, images, video and audio, browse the web, coordinate subagents and use computers across desktop, mobile and web environments. Its one-million-token context window is designed for large repositories, extended research tasks and workflows that must preserve state over many steps.
The weights are available through Amazon SageMaker JumpStart and several inference providers, including Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI. Meta also says the model is optimized to run on two Apple M5 Ultra Mac Studios through MLX, although that configuration remains far beyond ordinary consumer hardware.
The economics are part of the announcement. Together AI is offering free access through October 18 before moving to paid pricing, while Ollama has made the model available to several paid plans. That gives developers an immediate way to test whether its long-context and agent claims survive practical workloads rather than benchmark demonstrations.
Why it matters: Meta is making open weights a direct platform strategy, not merely a research gesture. A model this large is still expensive to operate, but broad provider support, local optimization and a million-token window could push coding agents toward longer, more autonomous sessions. The unresolved question is whether the active-parameter efficiency translates into dependable task completion outside curated evaluations.