Muse Spark Open: Meta's Open-Weight Model, Licence and How to Run It

Muse Spark Open is Meta's open-weight model, released on 15 October 2026: an 816-billion-parameter mixture-of-experts model with 42 billion parameters active per token, a 1M-token context window, and text, image, video and audio input. Meta calls it "the most capable US open-weight model to date" and built it for coding and agentic work. The weights are on Hugging Face under the Muse Model License, which is free for many developers but has conditions. You can run it yourself on very large hardware (Meta's showcase is two M5 Ultra Mac Studios) or call it through hosted APIs.

Key facts

What Muse Spark Open is

Meta describes it as the open-weight version of Muse Spark, its closed model line, with "additional safety post-training" because open weights can't be recalled once released. Meta's Muse Code product is powered by it.

The model card gives more architecture detail than the blog: a mixture-of-experts transformer with 62 layers, 256 experts with 8 routed per token, grouped-query attention, and a 1.8B-parameter Perception Encoder for images, video and audio. If "816B-A42B" in the name is unfamiliar, our mixture of experts entry explains it: you pay compute for 42B parameters per token, but you need memory for all 816B.

Reasoning effort takes low, medium, high or xhigh (default xhigh), and Meta recommends temperature 1.0 and top_p 1.0.

How it relates to Muse Wonder

This is the most common confusion, because the two launched in the same week.

So Muse Wonder is Meta's frontier tier that you rent from Meta; Muse Spark Open is the model you can own. Meta says Wonder uses "larger, stronger base models" than Muse Spark 1.3.

The licence: what you can and can't do

Muse Spark Open is an open-weight model, not open source in the OSI sense. From the Muse Model License summary on the model card:

• Attribution: you must display "Built with Muse Spark Open" in your product, docs or settings.
• Large hosts: a model-as-a-service provider with more than $20M aggregate gross revenue in any 12 consecutive months must notify Meta and sign a separate written agreement.
• Outputs: Meta claims no ownership of what the model generates.
• Usage policy: an Exhibit A policy prohibits uses such as violence, illegal activity, malware, deception and unlicensed professional practice.
• Age: not intended for users under 18.
• Gating: to download, you submit your full legal name, date of birth and organisation details on Hugging Face, and the form can't be edited afterwards.

For most companies building on it internally, these terms are workable. If you plan to resell it as an API at scale, read the full licence and talk to a lawyer.

How to run or access Muse Spark Open

Option 1: a hosted API (easiest)

Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Most expose an OpenAI-compatible endpoint, so you can try it without any hardware. Prices vary by provider; check each one.

Option 2: a server with vLLM or SGLang

The model card's quickstart for vLLM:

pip install vllm
vllm serve "meta-models/Muse-Spark-Open-816B-A42B"

That serves an OpenAI-compatible API on port 8000. SGLang works too, and Meta says the release ships with custom Helion GPU kernels in SGLang. Plan on at least 700GB of GPU memory for the NVFP4 weights.

Option 3: two Mac Studios with MLX

Meta's headline demo is two Apple M5 Ultra Mac Studios running MLX, each with up to 512GB of unified memory, which Meta says "behave like one supercomputer with a terabyte of unified memory". Meta calls it "the first model in its class to run on your desktop", but two top-spec workstations with a terabyte of memory between them is still a serious hardware investment.

Option 4: quantised GGUF with llama.cpp

Unsloth publishes GGUF builds and lists approximate total memory (RAM plus VRAM or unified memory) per bit width:

Unsloth says its 2-bit UD-IQ2_XXS fits on 256GB of RAM, and recommends UD-IQ2_S as the minimum for agentic tasks. Download and run:

hf download unsloth/Muse-Spark-Open-816B-A42B-GGUF \
    --local-dir unsloth/Muse-Spark-Open-816B-A42B-GGUF \
    --include "*UD-IQ2_XXS*"

./llama.cpp/llama-cli \
    --model unsloth/Muse-Spark-Open-816B-A42B-GGUF/UD-IQ2_XXS/Muse-Spark-Open-816B-A42B-UD-IQ2_XXS-00001-of-00006.gguf \
    --temp 1.0 --top-p 1.0

Unsloth notes that llama.cpp support may need its prebuilt fork until an upstream pull request is merged. The model card also lists quantised builds for Ollama, LM Studio and Jan. For choosing between those apps, see Ollama vs LM Studio, and for sizing a machine, our local LLM hardware guide.

Benchmarks (Meta-reported)

Meta's model card reports these at xhigh reasoning, among others: OSWorld Verified 80.4, MCP-Atlas 89.2, GPQA Diamond 89.0, Terminal-Bench 2.1 76.4, DeepSWE 1.1 60, and Artificial Analysis Intelligence Index 39. The card compares it with other open-weight models including Mistral Large 4, Kimi K3, GLM 5.3 and Qwen3.8, and it doesn't lead every row: it trails some of them on the Intelligence Index and HLE with tools. Meta says a full evaluation report is coming; treat these as vendor numbers until then.

Should you use it?

If you need a frontier-class model you control (data residency, air-gapped deployment, fine-tuning), Muse Spark Open is now one of the main candidates, alongside Mistral Large 4 once its weights land. If you just want to try it, use a hosted API first and only buy hardware after it passes your own tests. For a gentler start on local models, run an LLM locally with something smaller.