Meta releases Muse Spark Open: 816B open-weight model across 7 clouds on day one

#What happened

On October 15, 2026, Meta Superintelligence Labs released Muse Spark Open, an 816-billion-parameter mixture-of-experts model. Meta calls it "the most capable US open-weight model to date," a performance claim no independent benchmarks against rivals like DeepSeek, Qwen, or GLM have yet confirmed.

The model activates 42B parameters per token (256 experts per layer, 8 selected per token) and offers a 1-million-token context window. It accepts text, images, video, and audio as input and produces text output. Meta positions it for coding and agentic workloads and long-horizon tasks.

Distribution is broad on day one. Hosted inference is available from seven providers: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The weights can be downloaded through four channels: Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Microsoft confirmed support through its Azure AI Foundry blog, and Intel announced day-0 support.

Meta also demoed Muse Spark Open running on two Apple M5 Ultra Mac Studios via Apple's MLX framework.

Note that Muse Spark Open is not Meta's top model. The closed Muse Wonder sits above it. As Alexandr Wang put it, "Because of both safety and capabilities, we will generally not open-source our most powerful models."

#Why it matters

Day-one availability across seven inference providers and four weight channels makes this one of the widest open-weight launches from a US lab. Developers can pick a provider or self-host, which speeds up experimentation for coding agents and long-horizon tasks. The 1-million-token context window is aimed squarely at agentic workloads that need to hold large codebases or long documents in memory.

The release also draws a clear line in Meta's strategy. Muse Spark Open is strong, but the closed Muse Wonder remains the flagship. Open weights for a near-frontier model plus a stronger closed model mirrors the tiered approach other labs use.

Still, treat the headline performance claim as Meta's own for now. Until independent benchmarks against open-weight rivals are published, the "most capable US open-weight model" label is marketing rather than measurement.

Verified October 16, 2026.