Meta Muse Spark Open 的「開放權重」授權實際容許哪些用途

On October 15, 2026, Meta released the weights of Muse Spark Open, a natively multimodal mixture-of-experts model with 816B total parameters (42B active per token) and a 1M-token context window, built for coding and agentic work. Meta calls it "the most capable US open-weight model to date". That line is Meta's own claim, based on Meta's own benchmark table.

What happened

Weights are on Hugging Face (BF16 and NVFP4 checkpoints), plus Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
Hosted APIs launched the same day at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
It runs on SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth. Meta shows it running on two Apple M5 Ultra Mac Studios with 512 GB each.
Architecture (Hugging Face model card): 256 experts per MoE layer, 8 routed per token; text, image, audio and video in, text out.

Why people care

Chinese labs (Qwen, Kimi, GLM) have led open weights for most of 2026. This is a US lab answering with a large, agent-focused model that you can actually download and self-host.

What the coverage skips

1. Active vs total parameters. 42B active sets the compute per token. All 816B weights still have to sit in memory. The model card asks for roughly 700 GB of VRAM for NVFP4 and about 2 TB for BF16. Cheap per token is not cheap to host.

2. The license is not OSI open source. The Muse Model License opens like MIT ("free of charge... without restriction"), then adds conditions:

Section 2: any product that uses the model or a "Derivative" must prominently display "Built with Muse Spark Open" (UI, docs, website or settings).
"Derivatives" includes any model "trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof". A small model you fine-tune on Muse outputs counts.
Section 3: if you sell model access via an API ("Model as a Service") and your group's gross revenue passes $20 million in any 12 months, you must sign a separate agreement with Meta first. Internal use, research, and powering features of your own product are exempt.
The attached Usage Policy rules out, among other things, operating critical infrastructure, transportation technologies or heavy machinery, and says the model is not intended for anyone under 18.

3. The benchmarks are self-reported. Meta says a fuller evaluation report is coming. Independent indexes so far place it behind the strongest Chinese open models on general intelligence scores.

Facts vs prediction

Everything above comes from Meta's announcement, the Hugging Face model card and the license text. My read, a prediction rather than a fact: most teams will use it through hosted APIs, not local hardware, and the derivative clause will matter more than the $20M clause for small teams running distillation pipelines.

Practical takeaway

Read LICENSE.md before you build anything on it, especially if you plan to distill or fine-tune on its outputs.
Add the "Built with Muse Spark Open" line wherever you ship it.
Run 20 to 50 of your own real tasks through a hosted endpoint before trusting any leaderboard.

Sources:

Meta AI Research, Introducing Muse Spark Open (Oct 15, 2026): https://research.meta.ai/blog/muse-spark-open
Muse Model License: https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B/blob/main/LICENSE.md
Hugging Face launch post: https://huggingface.co/blog/meta-muse-spark-open

This post was written with AI assistance and edited for accuracy.