Muse Spark Open: hardware, licence and limits for product teams

Meta released the weights of Muse Spark Open on 15 October 2026. It is 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.¹ Meta calls it the most capable US open-weight model to date. For a product team the practical answer is that you can use it today through seven hosted APIs, self-hosting starts at roughly half a terabyte of accelerator memory, and the licence has two clauses worth reading before you build on it.

It shares its base model with Meta’s closed Muse Spark 1.3, with extra safety post-training because an open model runs without Meta’s system guardrails.²

This post covers what Meta published, what the hardware really is, what the scores do and don’t show, the licence terms that affect products, and how we would decide between the hosted APIs, our own GPUs and a closed model.

What Meta released on 15 October

Sources: the Hugging Face model card, Meta’s build guide and its deployment docs.³⁴⁵

The weights are on Hugging Face, SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hosted inference is on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.¹ Day-zero support covers SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, Unsloth and Apple’s MLX. PyTorch’s post adds Helion kernels that Meta reports run 1.03 to 3.95 times faster than hand-written CUDA kernels for operations like RoPE and RMSNorm on NVIDIA GB300, and a 3% end-to-end throughput gain on AMD MI350.² Those are vendor figures, and the 3% end-to-end number is the more useful one to plan with.

Two details from the deployment docs matter for anyone wiring it into an agent. Reasoning comes back in a separate reasoning_content field and tool calls in a standard tool_calls array once you start SGLang or vLLM with the muse parsers. In TensorRT-LLM, reasoning and the answer share one max_tokens budget, so a tight budget at xhigh effort can return empty content with finish_reason set to length.⁵ Budget for the reasoning tokens or lower the effort.

The hardware it actually needs

The headline that it runs on Macs is true, with a large asterisk. Meta’s recipe is two M5 Ultra Mac Studios with 512 GB of unified memory each, or four with 256 GB each, linked over Thunderbolt 5 with RDMA and running MLX tensor parallelism across the machines.⁴ Engadget reported that Apple has not announced price or ship date for the 512 GB configuration beyond “late October”, and estimated the four-machine 256 GB setup at roughly $38,000.⁶ Meta has not published a tokens-per-second figure for either Mac setup.

On GPUs, these are the configurations in Meta’s docs and the SGLang cookbook.⁵⁷

The model card puts the floor at 700 GB of VRAM for NVFP4 and 2 TB for BF16.³ The GGUF builds go much lower, but a 2-bit quantisation is a different model from the one Meta benchmarked. If you test one, run your evaluation on it directly instead of borrowing Meta’s scores.

The context window costs memory too

The 1M-token window is the other half of the memory bill. Using the model card’s shape, each token keeps 62 layers × 16 KV heads × 64 dimensions × 2 (keys and values) × 2 bytes in BF16. That is about 254 KB per token, or about 266 GB for one full 1,048,576-token sequence. This is our arithmetic and it is an upper bound: the SGLang recipes use an FP8 KV cache, which halves it, and set a sliding-window ratio flag, which suggests some layers keep only a window of recent tokens.⁷ The point stands for capacity planning. A server that holds the weights comfortably can still serve only a handful of long-context agent sessions at once, and TensorRT-LLM’s validated maximum sequence length is 49,152 tokens, not the full window.⁵

What the benchmark numbers say

Meta’s model card compares Muse Spark Open against several open models. Every number in it is vendor-reported, measured in each model’s own harness, and as of 19 October we have not found anyone who has reproduced them.³

The pattern is mixed. Muse Spark Open leads on MCP-Atlas and on Humanity’s Last Exam with tools, and trails Kimi K3 on the coding and terminal benchmarks that Meta’s own launch material emphasises. “Most capable US open-weight model” is a narrower claim than “most capable open-weight model”, and the table shows why the qualifier is there.

The one independent number so far is Artificial Analysis. On 19 October it listed Muse Spark Open at 39 on its Intelligence Index, 11th of 121 models, with a median output speed of 143.3 tokens per second and 0.79 seconds to first token across providers.⁸ Our earlier licence comparison noted that the highest-scoring open-weight model on the same index, Xiaomi’s MiMo-V2.6-Pro, sits at 46 under MIT (see which open source LLMs you can use commercially).

Meta also says the full safety evaluation report is “available soon”.¹ Until it is out, the release-threshold claim can’t be checked.

The licence terms that affect products

The Muse Model License is short. It grants use, modification, fine-tuning, distillation, merging, redistribution and sale, free of charge, subject to conditions.⁹ Three of them shape product decisions.

Attribution. You must “prominently display” the phrase “Built with Muse Spark Open” in the product’s interface, documentation, website or settings menu (section 2). A settings-menu notice satisfies it, so this is a small task. Put it on the launch checklist anyway.

Model as a service. This clause applies if you or your affiliates give third parties API access to a model and your combined gross revenue exceeds $20 million in any 12 months. You must then notify Meta and sign a separate agreement before offering Muse Spark Open or a derivative that way (section 3). Internal use is excluded, and so is using the model to power features of your own product that isn’t itself model access. A SaaS app with an AI feature is outside the clause. A platform that resells inference, or exposes a general model endpoint to customers, should get legal advice before it reaches that revenue.

Derivatives include models trained on outputs. Section 2 defines a derivative as any model “created, trained, modified (including fine-tuned), distilled” using the software or its outputs. Distil a smaller model from Muse Spark Open’s answers and that model carries the attribution and usage-policy obligations.

The usage policy attached as Exhibit A is where consumer products need care. Beyond the usual prohibitions, it says the model is “not intended for individuals under the age of 18”.⁹ If you build a learning app for schools or any product with teenage users, that line is a reason to pick a different model or to get a clear legal reading first. We read it as a statement of intended audience, and we are not lawyers.

The licence has no termination clause, no governing-law clause and no patent grant. Apache 2.0, which many other open models use, includes an explicit patent grant. For most product teams that difference won’t decide anything, but procurement teams at larger customers sometimes ask.

The Hugging Face download is gated behind a form asking for full legal name, date of birth and organisation details, and the form can’t be edited after you submit it.³ Fill it in once, as the organisation that will own the deployment.

Hosted, self-hosted or a closed model

Artificial Analysis lists a median hosted price of $1.25 per million input tokens and $4.25 per million output tokens across six providers, as of 19 October 2026.⁸ Meta’s docs give a 1M-token context on Fireworks and Baseten and 131,072 tokens on Groq, which Meta also describes as a limited preview. Together AI advertises a 99.9% SLA.⁵⁴ OpenRouter’s page says the model is not yet available through its EU in-region routing.¹⁰

The case for self-hosting is mostly about control: data residency, fine-tuning with TorchTitan, Fireworks or Together, and a model version that doesn’t change under you.²⁴ It is rarely about saving money at small volumes. Four Blackwell Ultra GPUs running all month cost the same whether they serve ten requests or ten million. Work out your monthly tokens, price them at the hosted rate, and compare that with your real GPU quote plus the engineering time to run it. Our prompt caching comparison is worth reading before that maths, because cached input changes the hosted side of it a lot.

For teams that worry about model churn, holding the exact weights is the strongest argument. You can keep the version you evaluated for as long as you like, where a closed model can be retired on a few months’ notice.

How we would evaluate it before building on it

1. Start on a hosted provider whose context limit matches your longest real prompt. Don’t benchmark on Groq at 131K if production needs 400K.
2. Run your own task set at high and xhigh effort. Record cost per task as well as accuracy, since effort changes output length.
3. Test tool calling through the same parser you will use in production (muse in SGLang or vLLM). Check that parallel calls and malformed arguments behave the way your agent expects.
4. If you might self-host, repeat the run on the NVFP4 build at the precision and context length you would actually deploy. Don’t trust a 2-bit GGUF to stand in for it.
5. Measure memory at your real concurrency and context, not just whether the weights load.
6. Read section 3 of the licence and the usage policy against your product’s audience and business model, and add the attribution notice to the launch checklist.

Steps 1 to 3 cost a few dollars of hosted tokens and a day of work. Only spend money on hardware after the model has passed them on your own tasks.

Frequently asked questions