Meta's 816B Muse Spark Open Runs on Two Mac Studios. Here's What It Really Takes

Meta has released the weights of Muse Spark Open, an 816-billion-parameter model it calls "the most capable US open-weight model to date." Anyone who accepts the licence can download it, fine-tune it and run it on their own hardware.

The headline demo is two Apple Mac Studios running the whole model on a desk, the result of an unusual Meta and Apple collaboration. There is a catch hiding in that demo: the 512GB memory option it relies on is, according to Apple's own store, still "coming late October."

Based on Meta's October 15 announcement, the official Hugging Face model card and Hugging Face's technical write-up, Artificial Analysis's independent scores, Apple's US store and reporting by 9to5Mac, checked October 16, 2026. We have not run the model. Benchmark figures are Meta's unless attributed to Artificial Analysis; memory arithmetic is ours.

What Meta released

According to Meta's announcement, Muse Spark Open is natively multimodal and handles up to 1 million tokens of context. It is a mixture-of-experts model: 816 billion parameters in total, but only about 42 billion are active for any given token.

Hugging Face's technical write-up fills in the design. Each layer has 256 experts and picks 8 per token. The important consequence for anyone hoping to run it at home is spelled out plainly: "All 256 experts stay in memory, so deployment needs the memory of an 816B-parameter model." Only a small slice does the work, but the whole model has to sit in memory.

The weights come in two formats, full-precision BF16 and a 4-bit NVFP4 version for Nvidia's newest GPUs. Meta lists downloads on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hosted access is offered by partners including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

How good is it?

Meta's model card reports, at its highest reasoning setting, 80.4 on OSWorld Verified, 76.4 on Terminal-Bench 2.1 and 89.0 on GPQA Diamond. These are Meta's own results.

The independent check comes from Artificial Analysis, which scores Muse Spark Open at 39 on its Intelligence Index. It ranks the model 10th of the 119 models in its comparison class, where the median score is 19. For context, Meta's closed models score higher on the same index: Muse Wonder Full at 56 and Muse Wonder at 52, as we covered in our Muse Wonder report. Meta is giving away a strong model, not its best one.

Artificial Analysis also measured hosted performance: about 102 tokens per second, with a typical price of $1.25 per million input tokens and $4.25 per million output tokens across providers. It estimates a cost of $0.84 per task in its index.

What it takes to run it yourself

Here is the memory arithmetic, which is ours. At 16 bits, 816 billion parameters need roughly 1.6 trillion bytes; Hugging Face puts the BF16 checkpoint at "about 1.5 TB." At 4 bits it is still roughly 400GB before any working memory. The model card says the NVFP4 version "requires a GPU with at least 700GB of VRAM," and Hugging Face says it fits a single Nvidia B200 server node.

Long conversations add more. Hugging Face estimates the BF16 key-value cache at about 266GB for a full 1-million-token context.

That is why the Mac route matters. Meta says Apple's M5 Ultra Mac Studio can give its GPU up to 512GB of unified memory with 1.2TB/s of bandwidth, so "two machines can hold an entire model," using Apple's MLX framework, 9to5Mac reported. Meta says it ran agentic coding and video-review tests entirely on-device this way.

Smaller setups are possible with heavier compression. Hugging Face reports a 3-bit MLX build that takes about 344GB and generates about 21 tokens per second on a single M3 Ultra Mac Studio with 512GB. A 2-bit llama.cpp build used 304GB and ran at 18 tokens per second. More aggressive compression usually costs some quality, and neither Meta nor Hugging Face published scores for these reduced builds.

What the licence allows

Muse Spark Open uses Meta's own Muse Model License, not a standard open-source licence. The terms on the model card are broader than many expected, with real conditions attached:

• Allowed: using, modifying, fine-tuning and distilling the model, and building derivative models. Meta claims no ownership of the model's outputs.
• The $20 million line: if a company offers the model as a hosted service and its gross revenue exceeds $20 million in any 12-month period, it must notify Meta and sign a separate agreement first. Internal use, research and embedding the model in your own product fall outside this rule.
• Attribution: products must prominently display "Built with Muse Spark Open."
• Usage policy: a list of banned uses such as malware, fraud and impersonation, and a statement that the model "is not intended for individuals under the age of 18."

Getting the files also means filling in a gated form on Hugging Face with your full legal name, date of birth and organisation details. The form warns it cannot be edited after submission.

The catch: the easy local setup isn't on sale yet

Meta's pitch leans on M5 Ultra Mac Studios with 512GB each. On October 16, Apple's US store said the "512GB memory option for M5 Ultra" is "coming late October" and showed no price for it. Two fully loaded machines will not be cheap, and you cannot order the exact configuration today.

Other open questions:

• Benchmarks are mostly Meta's. Artificial Analysis's 39 is the main independent number so far.
• Safety report pending. Meta says the model passed the thresholds its framework requires for release and that a full evaluation report will follow.
• Compression trade-offs are unmeasured. The single-Mac numbers above are speed figures, not quality scores.

In a sign of how unusual the partnership is, Meta AI chief Alexandr Wang told Axios he was "excited for mommy and daddy to be playing nice again," 9to5Mac reported. He credited former Apple executive Daniel Gross, now at Meta, with helping bring the companies together.

Who should care

For most people, the hosted API is the practical way to try Muse Spark Open, at a fraction of the hardware cost. Running it locally makes sense for teams that need data to stay on their own machines, want to fine-tune a large model, or are building a product that can live with the licence's attribution and $20 million terms.

If you want to start smaller, our beginner's guide to running AI locally covers models that fit on an ordinary laptop.

Frequently asked questions

Is Muse Spark Open really open source?

It is open-weight, not open source in the traditional sense. You can download, modify and distil it, but under Meta's own licence, which includes attribution, a usage policy and conditions for large hosted services.

Can I run it on my Mac?

Not on a typical one. The full model needs hundreds of gigabytes of memory. Meta's reference setup is two M5 Ultra Mac Studios with 512GB each, an option Apple lists as coming late October. Hugging Face ran heavily compressed versions on a single 512GB Mac Studio.

Is it better than ChatGPT or Claude?

Not on Artificial Analysis's index, where it scores 39, below Meta's own closed Muse Wonder models. Its appeal is that you can own and run it, not that it tops the charts.

By Meta's own account, Muse Spark Open is the most capable open-weight model a US company has released. Whether it lands on desks or stays in data centres will depend on how quickly those 512GB Mac Studios arrive, and what they cost when they do.