$ cat wiki/models/muse-spark-open.md
Muse Spark Open
Spec
| Attribute | Value |
|---|---|
| Developer | Meta AI |
| Released | 2026-10-15 |
| Announced | 2026-10-15 |
| Context window | 1M tokens |
| Pricing | unknown |
| License | Muse Model License; limited commercial terms |
| Availability | Downloadable weights and partner APIs |
| Meta describes a native multimodal model with 816B total parameters and 42B active, available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The announcement gives no API list price. Its license description is not a substitute for reading the complete terms. (source) |
Release Date
The weights and partner access were announced as available on 2026-10-15. Meta lists Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models as weight-distribution channels. (source)
Benchmarks
Meta calls this its most capable US open-weight release, but numerical capability charts are absent from the extracted article text. No independent rank or capability score is adopted here. Separately, Meta reports 1.03-3.95× kernel-level speedups on NVIDIA and AMD hardware and a 3% end-to-end speedup on AMD MI350 from Helion kernels in SGLang; these measure different scopes. (source)
Use Cases
The release targets coding, long-running agent tasks and multimodal reasoning. Meta demonstrates local Muse Code workflows on two Apple M5 Ultra Mac Studios using MLX, including iOS development and video editing. The article reports no token throughput or task-success rate for that setup. (source)
Compared To
Muse Wonder records a separate hosted agent offering. This release adds a downloadable deployment option; no numerical capability comparison with Wonder is established by the captured announcement. For Open-Weights Policy Fight, downloadable weights and unrestricted commercial use remain separate questions: Meta explicitly describes limited commercial terms. (source)
Sources
- Meta announcement and immutable snapshot. The linked model card, evaluation methodology and full license were not reviewed; Meta says a full evaluation report is forthcoming.