Meta’s Muse Spark Open is an open-weight AI model with 816 billion total parameters, of which 42 billion are active for each token, and Meta says it can run on your desk using two Mac Studios with the M5 Ultra chip. It is built for coding and agent-style work, accepts text, images, video and audio, and has a 1 million token context window. The weights went up on Hugging Face on October 15, 2026, and you can also try it through several cloud API providers without owning any hardware.
Apple did not co-build the model. Meta optimised it for Apple’s MLX framework and Mac Studio. Its two-Mac demonstration needs 512GB of memory in each machine, a configuration Apple says is coming in late October.
Muse Spark Open specs at a glance
| Total parameters | 816B (Hugging Face lists the release as 816B-A42B) |
| Active parameters | 42B per token (a mixture-of-experts design) |
| Context window | 1 million tokens |
| Input | Text, images, video and audio in; text out |
| Focus | Coding, agentic workflows, long multi-step tasks |
| Official weights | BF16 and NVFP4 on Hugging Face (gated, manual approval) |
| License | Muse Model License, custom, not an OSI open-source license |
| Released | October 15, 2026 |
Only 42B parameters are active for each token, but the full model still has to fit in memory. That makes memory capacity the first obstacle for running it locally; compute power and bandwidth still affect how fast it responds.
Meta calls it “the most capable US open-weight model to date” and publishes benchmark results on its model card. Those are Meta’s tests; independent results will tell us more about how it performs in everyday coding work.
What “open” means here
Muse Spark Open is open weight, not open source in the strict sense. You can download, run, fine-tune, distil and sell products built on it, subject to conditions in Meta’s custom license:
- You must show “Built with Muse Spark Open” in your product’s interface, documentation, website or settings.
- If you sell AI model access as a service (an API or similar) and your company and its affiliates together make more than $20 million in gross revenue over any consecutive 12 months, you need a separate agreement with Meta first. Internal use, research and using the model inside your own app’s features are exempt.
- An attached usage policy bans uses such as malware, weapons, fraud and impersonation, and says the model is not intended for people under 18.
- Downloading the official weights on Hugging Face requires a form with your legal name, date of birth, country and affiliation, and approval is manual.
Internal use and ordinary app features are outside that revenue clause. A company selling access to the model through an API should read the license before putting it into production.
How the two Mac Studio setup works
Meta’s argument is simple arithmetic. An M5 Ultra Mac Studio can be configured with 512GB of unified memory, and its GPU can use all of it. Two of them give you roughly 1TB, enough to hold the whole model. The M5 Ultra’s 1.2TB/s memory bandwidth is fast enough to read the 42B active parameters each token needs. Meta describes the pair as behaving “like one supercomputer with a terabyte of unified memory”, running through Apple’s MLX framework.
Why two and not one? The community MLX 4-bit conversion on Hugging Face (made by mlx-community, not Meta) keeps some weights in BF16 and comes to about 533GB with the vision encoder. That is more than a single 512GB Mac, so the model has to be split across two machines. Smaller 3-bit and mixed oQ quants also exist, but none of the official pages say they fit on one Mac or what quality they lose, and Meta has not published tokens-per-second figures for the Mac setup.
Meta says it tested agent workflows entirely on the two Macs with its Muse Code tool, including shipping bug fixes and features to the open-source Ice Cubes Mastodon app for iOS and driving DaVinci Resolve to pick and vertically crop the best moments from footage. Those are Meta’s own demos, not independent results.
What two M5 Ultra Mac Studios cost
This is where the “runs on your desktop” line gets expensive. Apple’s US store lists the Mac Studio with M5 Ultra from $5,499, and that base model has 96GB of memory. Meta’s setup needs 512GB, which requires the top 36-core CPU and 80-core GPU version of the chip. Apple’s configurator says the 512GB option is “coming late October” and has not listed its price yet.
Two base M5 Ultra machines cost at least $10,998 before tax in the US, but their 96GB memory configurations cannot run this model. The required 512GB upgrades will push the total higher; Apple has not announced their price. You also need a fast link between them, and both Macs have Thunderbolt 5 ports rated at up to 120Gb/s. Our NVIDIA PAIR home AI cluster guide covers another route to running large models locally, but neither setup is a casual desktop purchase.
How to try Muse Spark Open
No hardware: the easiest route is a hosted API. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights are also on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models if your company already uses one of those clouds.
Your own GPUs: Meta says it runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA hardware. It also ships custom Helion GPU kernels in SGLang, which Meta says speed up individual kernels by 1.03 to 3.95 times on NVIDIA and AMD GPUs.
On a Mac: Meta points to Apple’s MLX tools, and the community has converted the weights for MLX. Its 4-bit checkpoint is about 533GB, so the simple single-Mac example on the model page will not run on a 512GB Mac Studio. Meta says the model can run across two M5 Ultra Mac Studios, but it has not published a step-by-step two-machine setup in its announcement. If you need to try the model now, use a hosted provider rather than buying hardware on the strength of a demo.
Who Muse Spark Open is for
- Developers and teams who want a strong coding and agent model they can host privately, fine-tune with PyTorch’s TorchTitan and keep away from a third-party API.
- Companies with privacy rules that cannot send code or documents to a closed model, and already have GPU servers or a cloud account.
- Mac power users and studios who are already buying M5 Ultra machines for video or 3D work, for whom local AI becomes a bonus rather than the whole reason to spend.
- Not for most people. If you want a local assistant on a laptop or a Mac mini M6, this model is far too large. A smaller open model, or Muse Spark Open through a cheap hosted API, makes more sense.
My take
Muse Spark Open matters less because of the Mac demo and more because Meta is back to releasing a frontier-scale model whose weights developers can request. The license is friendlier than the headline “limited commercial terms” suggests, unless you are a large API seller.
The two Mac Studio claim is real in principle, but it is a showcase. It needs two of Apple’s most expensive desktops in a memory configuration you cannot order yet, and nobody has published independent speed numbers. I would try it through OpenRouter or another hosted provider first, see whether it beats the closed coding models you already pay for, and only then think about hardware.
