Meta Muse Spark Open Locally: 816B Weights, the Memory It Takes and the Licence

Meta released Muse Spark Open on 15 October 2026: an open-weight Mixture-of-Experts model with 816 billion parameters, a 1-million-token context, and text, image, video and audio input. Meta calls it “the most capable US open-weight model to date”. It is also one of the largest models anyone has tried to run at home.

This guide covers what the model is, the memory each version needs, which machines can actually hold it, and what the Muse Model License does and does not restrict. BFWR has not run the model. File sizes are read from the repositories, hardware fits are computed from our own database, and the only speed figures quoted are Hugging Face’s, attributed to them.

Quick answer

• Size: 816B parameters with 42B active per token, according to Meta. The full-precision weights are 1.64 TB, and Meta’s own 4-bit (NVFP4) version is 533 GB.
• Smallest download: unsloth’s 1-bit GGUF at 175.6 GB. The best-documented home setup is a 2- or 3-bit file of 300 to 370 GB on a 512GB Mac Studio.
• Software: llama.cpp and MLX support it in mainline releases. Ollama offers it only as a cloud model, and vLLM still needs a special Docker image.
• Licence: free to use, modify and sell, with an attribution line, a usage policy, and a separate agreement needed only for large “Model as a Service” businesses.
• The cheaper route: hosted APIs charge $1.25 per million input tokens and $4.25 per million output tokens on OpenRouter.

What Muse Spark Open is

Muse Spark Open comes from Meta Superintelligence Labs. Meta says it “is derived from the Muse Spark model we released in August, with additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.” The name causes confusion: Muse Spark 1.1, 1.2 and 1.3 are Meta’s hosted API models, and Muse Spark Open is the separate weights release.

From the model card:

• Architecture: 62 layers, 256 experts with 8 active per token, and grouped-query attention with 128 query heads and 16 key-value heads.
• Inputs: text, images, video and audio. Output is text only.
• Context: 1M tokens. Most layers use a 2,048-token sliding window, which vLLM says brings the cache at 1M tokens to “just over ¼” of what full attention would need.
• Reasoning: always on, with four effort levels from low to xhigh. The default is xhigh, and Meta’s llama.cpp notes say “Reasoning can’t be turned off.”

Hugging Face counts 822 billion parameters in the files rather than 816 billion. The difference is consistent with the vision tower and audio encoder, which Hugging Face counts and the “816B” label appears not to, though no source spells out Meta’s arithmetic.

How big are the files?

The model card states the requirement for its own formats bluntly: it “requires a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16.” That 700GB is for the 533 GB NVFP4 files plus cache and runtime, not a compressed size.

Every byte of these files has to be in memory, apart from a little headroom. Unlike DeepSeek V4.1-Flash-1017, which can leave its n-gram tables on the SSD, Muse Spark Open has no large part you can park on disk.

Which machines can hold it

Adding 10 GB to each file for the cache and runtime, and comparing that with the GPU-usable memory in the BFWR hardware database (on Macs, the default 75% of RAM):

• 1-bit (175.6 GB): a Mac Studio M5 Ultra with 256GB, or either 512GB Mac Studio.
• 2-bit (225.6 to 311.2 GB): only the 512GB Mac Studios, M5 Ultra or M3 Ultra.
• 3-bit (340 to 372 GB): the 512GB Mac Studios, with little room to spare at the default GPU limit. Our Mac VRAM optimizer explains how to raise it.
• 4-bit (470 GB and up): no single machine in the BFWR database. That is where Meta’s own setup comes in.

Unsloth itself says “The 2-bit UD-IQ2_XXS quant will fit on 256GB RAM”, which counts system RAM and GPU memory together, and it recommends UD-IQ2_S “as the minimum quant to use for agentic tasks”. That file is 275.5 GB.

Meta’s two-Mac setup

Meta’s route uses the 533 GB NVFP4 checkpoint split across machines. Its developer blog says to “Serve the model on either two Mac Studios, with M5 Ultra and 512GB of unified memory; or four Mac Studios with 256GB of unified memory, for a total of 1TB of unified memory”. The Mac cookbook adds that “The model doesn’t fit in one Mac’s memory”, that you need a direct Thunderbolt 5 link between every pair of Macs, and that “Four Macs serve fastest.” Apple sells the 512GB configuration only with the top 36-core CPU, 80-core GPU M5 Ultra.

Meta publishes no speed for this setup. The only Mac speeds we found are Hugging Face’s, on a single M3 Ultra with 512GB: a 2-bit llama.cpp quant “runs at 18 tokens/s” using 304 GB, and the oQ3e MLX quant ran “at about 225 tok/s prefill and 21 tok/s generation” in about 344 GB, with a short prompt. One 512GB Mac with a 2- or 3-bit file is therefore a far cheaper way in than two or four Macs with NVFP4, at some cost in quality.

Running it

• llama.cpp: support was merged on 15 October, and the first release with it is b11671. Meta’s blog gives llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S; change the tag to Q2_K for the 304 GB file.
• MLX: mlx-lm supports it from v0.32.0, and Meta’s multi-Mac guide uses mlx.launch with the NVFP4 checkpoint. On one Mac, the community MLX conversions are the practical choice, or oMLX for its oQ formats.
• Ollama: the library lists only muse-spark-open:cloud, which runs on Ollama’s servers. There is no local tag.
• vLLM: not in a release yet. vLLM publishes prebuilt muse-spark-open Docker images, and its support pull request was merged on 26 October. The latest release, v0.32.0 (18 October), predates the merge, so the Docker image or a nightly build is still the route.

Meta’s own repository is gated: you fill in a form, including your date of birth, before Hugging Face releases the files. The unsloth and ggml-org GGUF repositories are not gated, but the licence still applies to them.

The licence: what the $20 million line actually says

The Muse Model License opens like the MIT licence, granting the right “to use, copy, modify (including fine-tune), create derivative works, distill, merge, publish, distribute, sublicense, and/or sell copies of the Software”. Three conditions follow.

1. Attribution. You “shall prominently display “Built with Muse Spark Open” in the user interface of each product or service that uses the Software or any Derivatives, or in the accompanying documentation, website, or settings menu.” The licence defines Derivatives to include models trained using the Software “or any outputs or results thereof”, so a model distilled from its answers carries the same terms.

2. The usage policy. It is attached to the licence and binding. Among other things it states that “Muse Spark Open is not intended for individuals under the age of 18”, bans removing safety measures, and bans “Representing that the use of Muse Spark Open or outputs are human generated”.

3. The Model as a Service clause. This is the one widely described as a $20 million commercial limit, and that overstates it. A separate agreement with Meta is needed only when both of these are true:

• you or your affiliates provide third parties with access to AI models “via API or a similar interface” (the licence calls this “Model as a Service”);
• and “the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars” in any 12-month period.

The revenue test is the whole company group’s, not the model’s. The clause also expressly excludes “internal use, research, or any commercial activity that is not Model as a Service” and using the model “to power features of the Licensee’s own product or service”. So a large company that builds Muse Spark Open into its own app is outside the clause on this text, while a mid-sized API provider is inside it. Compare Krea 2, whose licence caps commercial use at US$1 million in revenue, or the attribution-plus-threshold terms on Qwen3.8-Flash-Next. This is a reading of the published text, not legal advice.

The licence also says “Meta claims no ownership of outputs generated by the Software.” There is no geographic exclusion: EU users simply contract with Meta’s Irish company.

Is it worth running locally?

For most people, no. A 256GB M5 Ultra Mac Studio was $9,499 in AppleInsider’s price guide on 11 October, and the 512GB model costs more. At $4.25 per million output tokens, $9,499 pays for about 2.2 billion output tokens from a hosted API, which can serve the 4-bit or full-precision weights instead of a 2-bit squeeze. The local case is the usual one: data that cannot leave your machine, offline use, or a Mac Studio you already own. Our API pricing calculator runs the comparison for your own token volumes.

Meta’s own benchmark table is worth reading before you commit. It scores Muse Spark Open at 39 on the Artificial Analysis Intelligence Index, against 45 for GLM 5.3 and 44 for Kimi K3 in the same table. “Most capable US open-weight model” is a claim about American labs, not about open models in general. For a model of similar ambition that is easier to run, see GLM-5.3-Flash.

FAQ

Can Muse Spark Open run on one computer?

Yes, at 1- to 3-bit, on a Mac Studio with 256GB or 512GB of memory. Meta’s own 4-bit NVFP4 version needs two 512GB Mac Studios or four 256GB ones.

Can I use it commercially?

Yes, with the “Built with Muse Spark Open” attribution and within the usage policy. Only a business selling model access through an API, with group revenue above $20 million a year, needs a separate agreement with Meta.

Is it on Ollama?

Only as a cloud model, muse-spark-open:cloud. To run it locally, use llama.cpp with a GGUF file or MLX on a Mac.

Why does Hugging Face say 822B?

It counts every tensor in the files, including the vision and audio encoders. Meta’s 816B figure appears to describe the language model alone.

Related guides and tools

• DeepSeek V4.1-Flash-1017 locally, the other giant open release of October
• Best Mac for local LLMs
• How to run GLM-5.3-Flash locally
• Open model directory, including the frontier models too large for the main list

Sources

• Muse Spark Open on Hugging Face (model card and LICENSE.md)
• Meta research blog, developer blog and Mac cookbook
• Hugging Face blog (Mac speed figures)
• vLLM blog
• unsloth GGUF files, unsloth guide and ggml-org GGUF files
• llama.cpp pull request #30616
• Ollama library tags and OpenRouter pricing

Updated 24 October 2026. Files, licence text, prices and support status were read on that date. Hardware fits are computed from the BFWR hardware database; speed figures are Hugging Face’s. Updated 29 October 2026: vLLM merged its support pull request on 26 October; it is not yet in a vLLM release.