Meta Muse Spark Open: 816 billion parameters you can download, under a license that isn't open source
Meta released Muse Spark Open on October 15, 2026, an open-weight mixture-of-experts model with about 816 billion total and 42 billion active parameters, a 1-million-token context and four input modalities. The weights are downloadable after a gated request, but the custom Muse Model License requires API resellers earning over $20 million in 12 months to sign a separate agreement with Meta
The three lines
- Scale — 816B total, ~42B active per token, 1M context, text/image/audio/video in
- Terms — custom license; $20M resale threshold and a mandatory 'Built with' label
- Standing — 39 on the Artificial Analysis index, 9th of 118 open-weight models above 150B
Key questions
- Muse Spark Open specs
- **An 816B-parameter MoE model with roughly 42B active per token.** | Item | Value | |---|---| | Total parameters | ~816B | | Active per token | ~42B | | Layers | 62 | | Experts | 8 of 256 per layer | | Context | 1,048,576 tokens | | Input / output | Text, image, audio, video / text |
- Muse Spark Open license commercial use
- **Free for most uses, with two strings attached.** | Term | Detail | |---|---| | API resale | Notify Meta and sign a deal above $20M revenue in 12 months | | Attribution | Display 'Built with Muse Spark Open' | | Allowed | Fine-tuning, distillation, merging, selling derivatives | | Access | Gated Hugging Face request |
- Can I run Muse Spark Open locally
- **Not on a normal PC.** | Setup | Memory | |---|---| | BF16 weights | ~1.5TB | | NVFP4 4-bit | 700GB+ GPU memory, one B200 node | | Mac demo | Two M5 Ultra Mac Studios, 512GB each |
Meta has handed out the weights to a model one step below its best. Muse Spark Open, released October 15, 2026, has about 816 billion parameters and, in Meta's words, is "the most capable US open-weight model to date." It is downloadable from Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry, and served by seven API hosts including Groq, Fireworks, Together AI and OpenRouter.
The word "open" needs care. The weights are public; the license is Meta's own, not Apache or MIT.
1. What was released
| Item | Value | Why it matters |
|---|---|---|
| Total parameters | ~816B | What you must store |
| Active per token | ~42B | What you actually compute |
| Layers / experts | 62 layers, 8 of 256 experts | About 3% of experts fire |
| Context | 1,048,576 tokens | Shared across all modalities |
| Runtimes | vLLM, SGLang, TensorRT-LLM, llama.cpp and more | AMD, Apple, Intel, Nvidia |
Compute is light for the size, but storage is not. The BF16 build is roughly 1.5TB; the NVFP4 4-bit build still needs 700GB-plus of GPU memory, or one Nvidia B200 node. Meta's smallest demo strings together two 512GB M5 Ultra Mac Studios. Our explainer "What is model quantization" walks through how 4-bit builds shrink models.
2. The license fine print
| Question | Muse Model License | MIT-style open source |
|---|---|---|
| Download weights | Yes, gated form (name, birth date, organisation) | Unrestricted |
| Fine-tune, distill, merge | Yes | Yes |
| Sell derivatives | Yes | Yes |
| Resell as an API | Separate deal above $20M revenue in 12 months | Unrestricted |
| Attribution | Prominent 'Built with Muse Spark Open' | Copyright notice |
Companies embedding the model in their own product are largely untouched. The clause bites cloud providers that would resell it wholesale, which reads as Meta declining to hand hyperscalers a free competing product.
3. How good is it
| Measure | Result | Source |
|---|---|---|
| Artificial Analysis index | 39; 9th of 118 open-weight models above 150B (median 19) | Independent |
| Closed sibling Muse Spark 1.3 | 48 | Independent |
| Output speed | 103.4 tokens/s | Independent |
| Hosted price | $1.25 in / $4.25 out per million tokens | Artificial Analysis |
| Terminal-Bench 2.1 / MCP-Atlas | 76.4 / 89.2 | Meta's own |
The same week, China's StepFun shipped Step 5 Preview (600B total, 27B active) and DeepSeek pushed a V4.1-Flash update, both open weights. Meta's release puts a US model back at the top end of that race, while keeping its strongest system, Muse Wonder announced October 13, closed.
4. What remains unconfirmed
- Meta's agentic benchmark scores await outside reproduction.
- There is no public evaluation of audio and video input yet.
- Meta's full safety evaluation is "coming soon"; once weights are out, they cannot be recalled.