Meta Releases Muse Spark Open: 816B Open-Weight Model, License and How to Run It
Meta released Muse Spark Open on 15 October 2026, an open-weight model from Meta Superintelligence Labs that the company calls "the most capable US open-weight model to date" (Meta AI Research). It is a large mixture-of-experts model built for coding and agentic work, and Meta says it is the first model in its class that can run on a desktop setup of two Apple Mac Studios.
What Meta released
According to Meta's announcement and the model card on Hugging Face, Muse Spark Open is a natively multimodal model that reads text, images, video and audio and writes text. Only a small slice of the model is active for each token, which keeps inference cheaper than the total size suggests.
| Detail | Muse Spark Open |
|---|---|
| Release date | 15 October 2026 |
| Parameters | 816B total, 42B active per token |
| Architecture | Mixture-of-experts (256 experts, 8 active) with a perception encoder |
| Context window | 1M tokens |
| Inputs / output | Text, image, video, audio in; text out |
| Checkpoints | BF16 (about 1.65 TB) and NVFP4 (about 537 GB) |
| License | Muse Model License |
The checkpoint sizes come from Meta's developer docs. The model card lists a reasoning-effort setting (low, medium, high or xhigh, with xhigh as the default) and recommends high or xhigh for coding and agent tasks.
How it compares (Meta's own numbers)
Meta's model card compares Muse Spark Open with other large open-weight models. These are vendor-reported benchmarks, so treat them as Meta's claims until independent tests catch up. On Meta's table, it scores 39 on the Artificial Analysis Intelligence Index, 80.4 on OSWorld Verified and 76.4 on Terminal-Bench 2.1, while some rivals in the same table, such as GLM 5.3 and Kimi K3, score higher on several coding benchmarks. The Decoder framed the release as Meta's answer to Chinese open-weight leaders like Qwen, GLM and Kimi.
The license: open, with conditions
Muse Spark Open is not under Apache 2.0 or MIT. The Muse Model License lets you use, modify, fine-tune, distill and sell products built on the model for free, with a few conditions worth reading before you build:
- Attribution: products using the model or a derivative must prominently show "Built with Muse Spark Open" in the interface, docs, website or settings.
- Large API providers need a deal: if you offer AI models to others as an API service ("Model as a Service") and your group's revenue is over 20 million US dollars in any 12-month period, you must notify Meta and sign a separate agreement. Internal use, research and using the model to power features in your own non-API product are excluded from this rule.
- Usage policy: an acceptable-use policy applies, and Meta says the model is not intended for people under 18.
Can you run it yourself?
Not on a normal laptop. The model card says self-hosting needs at least about 700 GB of GPU memory for NVFP4 and about 2 TB for BF16. Meta's docs give serving recipes for 4 NVIDIA B300 GPUs (NVFP4) or 8 B300 GPUs (BF16), and Meta says two Mac Studios with M5 Ultra chips (512 GB of unified memory each) can run it with Apple's MLX framework. Supported engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp and Unsloth (Meta AI Research).
For most builders, the practical route is a hosted API. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as API providers at launch.
What it means for builders
- Another strong open option for agents. If you build coding or tool-using agents and want weights you can host, fine-tune or move between providers, this joins Reflection's Beam and NVIDIA's Nemotron 3.5 from the same week (see our 18 October roundup).
- Try before you commit. Test it through an API provider such as OpenRouter, Together AI, Groq or Fireworks AI before planning any self-hosting.
- Check the license for SaaS plans. The attribution line and the 20-million-dollar API-provider clause matter if you resell model access.
- Want something that runs locally today? Our local LLM tools list covers smaller open models you can run on one machine, and our free LLM APIs list covers free tiers for testing.
Sources
- Meta AI Research: Introducing Muse Spark Open
- Hugging Face: Muse Spark Open model card and Muse Model License
- Meta developer docs: Get the model
- TechCrunch: Meta launches an open-weight AI model that runs on Macs
- The Decoder: Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi
FAQ
What is Muse Spark Open?
It is an open-weight AI model released by Meta on 15 October 2026. It has 816 billion total parameters with 42 billion active per token, a 1M-token context window, and accepts text, image, video and audio input. Meta built it for coding and agentic work.
Is Muse Spark Open free for commercial use?
Mostly, under the Muse Model License. You can use and sell products built on it for free, but you must display "Built with Muse Spark Open", and companies offering models as an API service with over 20 million US dollars in yearly revenue must sign a separate agreement with Meta.
What hardware do I need to run it?
According to Meta, at least about 700 GB of GPU memory for the NVFP4 version (for example 4 NVIDIA B300 GPUs) or about 2 TB for BF16. Meta also says it runs on two Mac Studios with M5 Ultra chips using MLX.
Where can I try it without my own GPUs?
Meta lists API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.