Meta just did something it hasn't done in over two years: it released the weights of a frontier-class model. Muse Spark Open, launched October 15, 2026 by Meta Superintelligence Labs, is an 816-billion-parameter mixture-of-experts model — but only about 42 billion parameters fire per token, which is what keeps it fast and affordable to serve.
On the independent Artificial Analysis Intelligence Index it scores 39 — the highest of any US-built open-weight model, though still behind China's GLM-5.3 (45) and Kimi K3 (44). It is built for coding and agentic work: multi-step reasoning, tool use, and a 1M-token context window. Hosted APIs (Together AI, Fireworks AI, Groq, and four others) price it at a median of $1.25 per million input tokens and $4.25 per million output tokens.
The model ships under Meta's custom Muse Model License, not Apache 2.0 or MIT. Two terms matter:
Self-hosting is mostly theoretical: the NVFP4 weights need roughly 700 GB of VRAM, and Meta's own reference setup is two 512 GB M5 Ultra Mac Studios. For nearly everyone, the hosted API is the sensible route.
Spark Open is the previous generation of Meta's model line released as free weights — weaker than the closed Muse Wonder (52 on the index), but downloadable, fine-tunable, and yours to run. If you build agents, it is worth a one-week test through a hosted provider.
This is a condensed developer explainer. The full breakdown — complete specs, pricing tables, the license terms in plain English, and head-to-head comparisons with Muse Wonder, DeepSeek, and Kimi — is on NeurAITool: Muse Spark Open: Meta's 816B Open-Weight Giant Explained.