Muse Spark Open
AvailableMuse Spark Open (Muse-Spark-Open-816B-A42B) is Meta's open-weight Muse Spark model, released 2026-10-15 and billed by Meta as the most capable US open-weight model to date. Mixture-of-Experts, 816B total / 42B active, 1M-token context, natively multimodal (text, image, audio and video in; text out). Released under the Muse Model License: free for internal use and in-product use; companies over $20M revenue must sign a separate agreement to offer it as a model-as-a-service; products must display "Built with Muse Spark Open"; military / critical-infrastructure use barred (with a US-government carve-out). Weights on Hugging Face, SageMaker, Vertex Model Garden and Microsoft Foundry; hosted by Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together (~$1.25 / $4.25 per 1M). Meta reports wins over Nemotron 3 Ultra and Thinking Machines' Inkling, leads over Kimi K3 / GLM 5.3 / Qwen3.8 on agentic benchmarks (JobBench, MCP-Atlas), but trails on coding (Terminal-Bench 2.1 76.4 vs Kimi K3 85.0). Appears to be the long-promised open-weight release of the Muse Spark line.
Specifications
- License
- Open weights · Muse Model License
- Weights
- Downloadable
- Architecture
- Mixture-of-Experts
- Parameters
- 816B · 42B active
- Context window
- 1.0M tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- $1.25 / $4.25
- Modalities
- TextVisionAudioVideoCode
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.