Muse Spark Open
Meta's 816B open-weight agentic coding model — 42B active params, natively multimodal, 1M context, tuned to run on Mac Studios.
About
Muse Spark Open is the open-weight release of Meta's Muse Spark model, published October 15, 2026. It is a mixture-of-experts model with 816B total parameters and 42B active per token (256 experts, 8 routed per token), natively multimodal across text, image, video, and audio, with a 1M-token context window. Meta built it for coding and long-horizon agentic work: the announcement demos include shipping real iOS changes to an open-source Mastodon client and driving DaVinci Resolve frame-by-frame through Muse Code. In an unusual Meta-Apple collaboration, the model is optimized for Apple's MLX framework so it runs on two to four M5 Ultra Mac Studios (1TB pooled unified memory) as a desktop setup. Weights ship on Hugging Face in BF16 and NVFP4, and hosted APIs are available through Together, Fireworks, Groq, OpenRouter, and others. The Muse Model License requires attribution ('Built with Muse Spark Open') and a separate agreement for model-as-a-service businesses earning over $20M a year.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Artificial Analysis Intelligence Index 39 — reported as the highest of any non-Chinese open model (vendor-reported). Meta's own table claims leads on JobBench, MCP-Atlas, HLE-Diamond, and ZeroBench, but trails Kimi K3 and Qwen3.8 on Terminal-Bench 2.1 and DeepSWE. Independent replications pending.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 5/10
Scored 8/10 (METRICS tier, metrics-v1). Evidence: Meta is a frontier lab shipping a genuinely open 816B-weight release with published license terms, Hugging Face distribution in two precisions, and a multi-vendor hosted API ecosystem on day one; the Apple MLX collaboration is a concrete, verifiable deployment story. Docked for: (1) the 'most capable US open-weight' framing is vendor-reported and Meta's own table shows it trailing Kimi K3 and Qwen3.8 on coding benchmarks, with no independent replication yet; (2) the Muse Model License restricts commercial API hosting above $20M revenue and mandates attribution, so it is open-weights, not open-source; (3) practical self-hosting needs ~700GB VRAM (NVFP4) or two high-end Mac Studios — out of reach for most individuals. License: Muse Model License (custom, limited commercial).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Meta frontier lab; 816B open-weight release Oct 15 2026
- · HF distribution (BF16 + NVFP4) + hosted APIs via Together, Fireworks, Groq, OpenRouter on day one
- · Meta-Apple MLX collaboration for Mac Studio deployment is concrete and verifiable
- · AA Intelligence Index 39 reported as top non-Chinese open model (vendor-reported)
- · 'Most capable' framing is vendor marketing; Meta's own table trails Kimi K3/Qwen3.8 on coding
- · manual review: no independent benchmark replication yet; license restricts >$20M/yr API hosting