Muse Spark Open
Muse Spark Open is Meta's open-weight Muse model, released October 15, 2026 under the Muse Model License, which carries limited commercial terms.
Provider
Meta
Model family
Meta Muse
Open-weight multimodal MoE LLM
Cost tier
Spark Open
Status
Current
Release Oct 15, 2026
Technical specs
- Inputs
- text, image, audio, video
- Outputs
- text
- License
- Muse Model License (limited commercial terms)
- Weights
- meta-models/Muse-Spark-Open-816B-A42B, NVFP4 variant
- Total Params
- 816B
- Active Params
- 42B
- Context tokens
- 1,000,000
- Capabilities
- Open weights on Hugging Face, including an NVFP4 variant
- 816B total / 42B active Mixture-of-Experts
- 1 million token context window
- Text, image, audio, and video input
Why teams choose it
Complex reasoning
Useful for workflows that require structured thinking, multi-step logic, and deeper analysis than lightweight models provide.
Long-context analysis
Helps teams summarize, compare, and extract insights from long documents without losing important nuance.
Meta roadmap vigilance
Use published model pages—not stale marketing blurbs—for modalities, quotas, pricing, and policy; schedule revalidation tied to vendor release notes.
Cost-efficient routing
Useful as part of a routing stack where cheap models handle drafts and confirmations and this tier handles genuinely hard passages.
Tradeoffs to know
- The Muse Model License has limited commercial terms; read it before production use.
- Self-hosting an 816B MoE needs multi-GPU infrastructure even with NVFP4.
- Not the same checkpoint as hosted muse-spark-1.3.
Benchmarks
No independent or vendor benchmark scores in this catalog entry.
Meta Muse family lineup
Related models
Muse Spark Open FAQ
What is Muse Spark Open?
Muse Spark Open is Meta's open-weight Muse model, released October 15, 2026 under the Muse Model License, which carries limited commercial terms. It is a Mixture-of-Experts model with 816 billion total and 42 billion active parameters, a 1 million token context window, and text, image, audio, and video input with text output. Weights are on Hugging Face as meta-models/Muse-Spark-Open-816B-A42B, with an NVFP4 quantized variant. It is a separate open release, not the hosted Muse Spark 1.3 API model.
When does Muse Spark Open fit best?
Self-hosted multimodal assistants with long context
What should teams watch out for with Muse Spark Open?
The Muse Model License has limited commercial terms; read it before production use.
Explore next
Models, tools, and comparisons that connect to this reference.