Meta Muse Spark Open: 816B Weights, One Catch

Meta released the weights for Muse Spark Open on 15 October 2026: 816 billion total parameters, about 42 billion active per token, a one million token context window, and text, image, video and audio in with text out. You can download it today. Whether you can ship a product on it depends on two clauses in the licence file rather than anything in the announcement, and that is the part worth your attention.

What Meta actually shipped

The architecture is a sparse mixture-of-experts transformer with a perception encoder on the front. The model card lists 62 layers, 256 experts with 8 activated per token, 128 query heads against 16 key-value heads, and a vocabulary of 202,048.

Meta describes it as "trained for long-horizon agentic work such as coding and customer service agents". That framing matters more than the parameter count. The pitch is an agent that reads a large codebase, plans a change, calls tools, checks the result and corrects itself when a step fails, rather than a model that writes a function on request.

Weights are on Hugging Face, and seven hosted providers had it at launch, including Baseten, Fireworks AI, Modal, Nebius, OpenRouter and Together AI, with Groq in limited preview. To run it yourself, Meta's launch guide names SGLang, vLLM, TensorRT-LLM, MLX, llama.cpp and Unsloth, verified across NVIDIA, AMD, Intel and Apple silicon.

The two clauses that decide whether you can use it

The weights are downloadable. "Open" in the name is doing some work.

First, the revenue clause. If you or your affiliates offer AI models to third parties as a service, and your aggregate gross revenue passes 20 million US dollars in any consecutive twelve month period, you have to notify Meta and sign a separate written agreement before offering Muse Spark Open as part of that service. Read the carve-outs carefully, because they are generous: internal use, research, and using the model as a component of your own product are all excluded. A founder whose app happens to call the model is not the target. A company reselling inference is.

Second, the attribution clause. You must prominently display "Built with Muse Spark Open" in each product, documentation, website or settings menu that uses the model or anything derived from it. That is a design decision, not a legal footnote. Decide now where that string lives in your interface, because retrofitting it into a settings page later tends to be the thing nobody owns.

The licence also carries a usage policy in Exhibit A covering the usual prohibited categories, and states the model is not intended for people under 18. Our guide to what different AI model licences let you ship covers how these clauses compare across licence families.

The benchmark table is on the card, not in the announcement

Worth noting for anyone evaluating this: the launch post publishes no benchmark results. The numbers live on the model card, and they are Meta's own, run at what it calls xhigh reasoning.

Self-reported numbers from a model vendor are a starting point, not a verdict, and sorting the signal from the noise is most of keeping up with AI news. The agentic scores are the interesting ones here because they match the stated purpose, and because third-party reproduction is what separates a release that changes your stack from one that changes your feed.

The launch page does carry third-party figures, on serving speed rather than quality. RadixArk reports 3.7 times single-request throughput on AMD MI355X hardware using speculative decoding, and 2.9 times at sixteen concurrent requests.

Can you run Muse Spark Open on your own hardware

Mostly no, and the two Meta pages do not quite agree, which is instructive.

The launch post puts BF16 weights at roughly 1.6 TB and the NVFP4 checkpoint at about 530 GB, served on four Blackwell Ultra GPUs or eight H200s. The model card asks for at least 700 GB of VRAM for NVFP4 and 2 TB for BF16. Both can be true: weight size is not the memory a server needs once the key-value cache for a long context is resident. Plan against the higher figure.

The desktop claim is the one that travelled. Meta says two M5 Ultra Mac Studios with 512 GB each can hold the NVFP4 model over Thunderbolt 5, and the smallest Unsloth build, around two bits, fits in roughly 300 GB of combined RAM and VRAM. Genuinely new for a model this size, and still a five figure hardware bill before you serve a token. If you are sizing this up, start with how much VRAM a local model actually needs, then whether self-hosting beats an API at your volume. For most teams at most volumes the answer stays the API.

What to do with this

If you build agents, test the long-horizon behaviour rather than the leaderboard position: give it a multi-file change in your own repository, watch whether it checks its own work, and compare cost per completed task against what you run now. The mixture-of-experts design is why a model this large can be priced like a much smaller one, since only about 42 billion parameters fire per token.

If you resell inference, read the revenue clause first. If you ship a product, decide where the attribution string goes. The rest can wait for someone else's reproduction.