Muse Spark Open: what Meta's open-weight license actually lets you do

Meta released Muse Spark Open's weights on Oct 15, 2026 (816B total, 42B active, 1M context). Five slides on what shipped, the three license clauses (credit line, derivatives that reach outputs, the $20M API clause), and what to check before you build. Sources: https://research.meta.ai/blog/muse-spark-open and https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md

Transcript

1. Muse Spark Open What Meta's open-weight license actually lets you do Weights released October 15, 2026 Sources: Meta AI Research, Hugging Face LICENSE.md Clydara 1/5
2. What shipped 816B total parameters, 42B active, 1M-token context Weights: Hugging Face (BF16 + NVFP4), SageMaker, Model Garden, Foundry Hosted APIs: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together Runs on SGLang, vLLM, TensorRT-LLM, MLX, llama.cpp, Unsloth Model card: ~700 GB VRAM for NVFP4, ~2 TB for BF16 Clydara 2/5
3. The license, in three clauses Reads like MIT, then adds conditions Show "Built with Muse Spark Open" in any product using it Derivatives include models trained or distilled on its outputs API businesses over $20M revenue in 12 months need a Meta agreement Internal use, research and your own product features are exempt Clydara 3/5
4. What coverage skips 42B active is compute per token; all 816B still sit in memory Not an OSI open source license Benchmarks are self-reported; full evaluation report pending Clydara 4/5
5. Before you build Read LICENSE.md, especially before any distillation pipeline Add the credit line wherever you ship it Run 20 to 50 of your own tasks on a hosted endpoint first Full write-up: clydara.wordpress.com Clydara 5/5