Muse Spark Open: can you run Meta’s open-weight model, and Muse Code, locally?
Meta released Muse Spark Open on October 15: 816B parameters, a custom license, and a hardware bill most developers can’t meet. Here’s what the license allows, what it takes to run, and what Meta has and hasn’t said about running Muse Code on it.
What Meta released
On October 15, 2026, Meta published the weights of Muse Spark Open, a large open-weight model from the Muse Spark family. Its main specs:
- Size: 816 billion total parameters, 42 billion active per token (a mixture-of-experts design).
- Context: 1 million tokens.
- Input: text, images, video, and audio. Output: text.
This is a different kind of release from Muse Glimmer, the 30B open model Meta shipped on August 10. Glimmer is meant to run on a single consumer GPU. Spark Open is meant to run in a data center or on a serious workstation cluster.
The license in plain English
Spark Open ships under the Muse Model License, a custom license written by Meta. It is not an OSI-approved open-source license, so “open-weight” is the accurate term. In plain terms:
- You may run, modify, fine-tune, distill, and redistribute the model, provided you follow the license conditions and Meta’s usage policy.
- You must show “Built with Muse Spark Open” in products that use it.
- If you are a large model host — a company with more than $20 million in gross revenue over 12 months from offering the model as a model-serving business — you must notify Meta and sign a separate agreement.
- Your outputs are yours. Meta claims no ownership of what the model generates.
For most teams the revenue clause won’t apply. Hosting providers and resellers should read it in full. Glimmer, by contrast, is Apache 2.0, with no attribution label and no revenue threshold.
What hardware does it actually need?
A lot. Meta’s guidance is roughly 700 GB of accelerator memory in NVFP4 precision and about 2 TB in BF16. The 42B active parameters keep each token relatively cheap to compute, but all 816B parameters still have to sit in memory.
Meta’s own demo used two M5 Ultra Mac Studios — up to 512 GB of memory each — running the model with MLX. That is the most accessible setup Meta has shown, and it still takes two top-end machines. Spark Open is not a laptop model, and a single high-end gaming GPU won’t run it.
Where to get it
- Weights: Hugging Face (gated,
meta-models/Muse-Spark-Open-816B-A42B), Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry. - Inference engines: SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth. Meta points to TorchTitan for fine-tuning.
- Hosted APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Together lists it at $1.25 input, $0.15 cached input, and $4.25 output per 1M tokens. Groq caps the context at 131K tokens instead of the full 1M.
Together’s price is the same as Meta’s own rate for hosted muse-spark-1.3. If all you want is per-token access, open weights don’t make it cheaper. They pay off when you need to self-host, fine-tune, or keep data on your own hardware.
How does it score?
Meta’s launch table, with the usual caveat that these are the vendor’s own numbers:
| Benchmark | Muse Spark Open | Note from Meta’s table |
|---|---|---|
| Terminal-Bench 2.1 | 76.4 | Kimi K3 leads at 85.0 |
| MCP-Atlas | 89.2 | Leads the comparison |
| JobBench | 61.8 | Leads the comparison |
| Artificial Analysis Intelligence Index | 39 | Below GLM-5.3 (45) and MiMo-V2.6-Pro (46) |
Spark Open looks strongest at tool use and agent-style tasks, and weaker on overall intelligence measures than the best rival open models. For comparison, VentureBeat put hosted Spark 1.3 at 61 on the same index, as noted in our Spark 1.3 write-up. The open model is not a free copy of the flagship.
Can you run Muse Code locally on it?
Partly — on Meta’s word, but not yet with Meta’s instructions. Meta says Muse Code runs on Spark Open for on-device agentic workflows on Apple hardware. Its demos showed the agent editing the Ice Cubes Mastodon app and operating DaVinci Resolve.
What Meta has not published is a step-by-step setup: there is no documented local-provider configuration that tells you how to point Muse Code at a self-hosted Spark Open endpoint. CLI 1.4.2 added --provider and --model flags to muse serve, but Meta doesn’t document them as a route to a local Spark Open server. So:
- Fully offline Muse Code is something Meta has demonstrated, not documented. Expect to experiment, and don’t plan a team rollout around it yet.
- Hardware is the real limit. The documented demo needed two high-memory Mac Studios.
- Subscription models don’t come with it. Running locally doesn’t give you Muse Wonder or Wonder Full; those are hosted only.
If that changes, it will show up in the docs and our changelog.
Spark Open vs. Muse Glimmer vs. hosted Spark 1.3
| Muse Spark Open | Muse Glimmer | Muse Spark 1.3 (hosted) | |
|---|---|---|---|
| Released | October 15, 2026 | August 10, 2026 | September 2, 2026 |
| Size | 816B total, 42B active | 30B | — |
| License | Muse Model License (custom) | Apache 2.0 | Proprietary, API and Meta products only |
| Hardware | ~700 GB (NVFP4) to ~2 TB (BF16) | Single consumer GPU, ~20 GB quantized | None — Meta hosts it |
| Context | 1M tokens | — | 1M tokens |
| Price | Your hardware, or $1.25 / $4.25 per 1M on Together AI | Your hardware | $1.25 / $4.25 per 1M (standard) |
| Muse Code | Meta says it runs on Apple hardware; no documented setup | Not documented for Muse Code | The default model |
Who is it for?
- Enterprises and labs that must keep data on their own infrastructure, or want to fine-tune or distill a large agentic model.
- Inference providers, which explains the long list of launch-day hosts — and the revenue clause aimed at them.
- Owners of very large Apple or GPU setups who want to experiment with a local agent ahead of official documentation.
For everyone else, the practical choices haven’t changed: hosted Spark 1.3 through Muse Code or the Meta Model API if you want the agent, and Glimmer if you want something that runs on one GPU. Our comparison page shows how all of these stack up against rival agents.