Meta's Muse Spark Open: Running an 816B Model on Mac Studios

By Jaspal Singh October 16, 2026 Updated: October 16, 2026

Muse Spark Open is Meta's new open-weight AI model: 816 billion parameters, of which about 42 billion work on each token, with a one-million-token context window. Two Apple M5 Ultra Mac Studios linked by Thunderbolt 5 run it at about 25 tokens per second, putting a frontier-class coding model on a desk.

What Happened

On October 15, 2026, Meta published the weights of Muse Spark Open and described it in its announcement on the Meta AI research blog as "the most capable US open-weight model to date." The model is aimed at coding and agentic work: long-horizon tasks, multi-step reasoning and complex software engineering. It accepts text, images, video and audio as input and produces text.

The weights are on Hugging Face in BF16 and NVFP4 formats, and through Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is available from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and Meta lists support in SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth.

The unusual part is Apple. The same day, Apple's machine learning team posted benchmarks for running Muse Spark Open across several Mac Studios with MLX, its open-source framework for Apple silicon. Meta calls the pairing "the first model in its class to run on your desktop." Meta AI chief Alexandr Wang told 9to5Mac the two companies had been working together for several months, a notable thaw after years of public friction over privacy and app store rules.

Bloomberg framed the release as Meta opening one model while keeping its best technology closed. The proprietary Muse Spark line, which Meta Superintelligence Labs launched in April to power Meta AI, is still not downloadable.

Why It Matters

Until now, the strongest open-weight models have mostly come from Chinese labs such as Moonshot (Kimi), Z.ai (GLM) and Alibaba (Qwen). Muse Spark Open gives companies that will not deploy Chinese-built weights, for compliance or procurement reasons, a large US model they can host themselves.

The Mac angle matters for a different group. A model of this size has normally needed a server with eight data-centre GPUs. Apple's unified memory lets the M5 Ultra's GPU address up to 512 GB per machine, so two Mac Studios hold the compressed weights between them. For a team that does not want code leaving the building, that is a very different purchase from a GPU server.

Strategically, it is a middle path for Meta: developer goodwill from a strong open model, without giving away the flagship that runs inside WhatsApp and Instagram.

How It Works

A mixture of experts, not one giant network

According to the Muse Spark Open model card on Hugging Face, the model is a mixture-of-experts transformer with 62 layers and 256 experts per layer. For each token, a router picks 8 of those 256 experts, so only about 42 billion of the 816 billion parameters do work at any moment. A separate 1.8-billion-parameter Perception Encoder turns images and video frames into tokens the language model can read.

This design is why a desktop can run it at all. Generating text is mostly limited by how fast a chip can read weights from memory. A dense 816B model would have to read every parameter for every token; Muse Spark Open reads only the slice its router chose. Meta's blog puts it plainly: the M5 Ultra's 1.2 TB/s of memory bandwidth is "fast enough to read the small slice each token needs." All 816 billion parameters still have to sit in memory, though, because the next token may need a different set of experts.

How much memory it needs

Apple's post sizes the model at about 1.6 TB in BF16 and about 493 GB in NVFP4, a 4-bit format. The model card is more conservative and recommends at least 700 GB of GPU memory for NVFP4 and 2 TB for BF16, leaving room for the key-value cache that grows with long prompts. Either way, one 512 GB Mac Studio is not enough for the 4-bit weights plus working memory. Two are: Meta says the pair "behave like one supercomputer with a terabyte of unified memory."

Splitting one model across several Macs

Apple tested two ways of dividing the model, and the difference explains the speed figures:

Pipeline parallelism splits the model by depth. The first Mac runs the first half of the layers and hands its result to the second. It needs little communication, but each token still passes through both machines in turn, so generation runs at roughly single-machine speed.
Tensor parallelism splits every layer by width, so each Mac holds part of every layer and both work on each token at once. That requires many small exchanges per layer, which only works because macOS 26.2 added RDMA (remote direct memory access) over Thunderbolt 5, letting one Mac write straight into another's memory with very low latency. MLX 0.30.1 and later build on that with a communication backend called JACCL.

Apple's measurements used the 4-bit model with an 8,192-token prompt, generating 128 tokens. Two Macs with pipeline parallelism produced 15.74 tokens per second; two Macs with tensor parallelism reached 24.99; four Macs with tensor parallelism reached 36.73, which Apple notes is well above average human reading speed. Time to first token on four machines was 4.8 seconds. Apple credits the M5 GPU's Neural Accelerators for up to 4.6 times faster time to first token than the M3 Ultra.

In practice, running it is a short list of commands: mlx.launch with a host file naming each Mac, the JACCL backend, and then mlx_lm.server to expose an OpenAI-style local endpoint that coding tools can call. Apple's post pairs it with the OpenCode agent; Meta's own demos used Muse Code, its terminal coding tool, to ship bug fixes and new features to the Ice Cubes Mastodon app for iOS and to drive DaVinci Resolve to review and crop video footage, all on-device.

How it scores

Meta's model card, as relayed in launch coverage, reports 89.2 on MCP-Atlas (tool use) and 61.8 on JobBench, the highest of the models Meta compared. On coding it is behind the best Chinese open models: 76.4 on Terminal-Bench 2.1 against 85.0 for Kimi K3 and 83.9 for GLM 5.3. On GPQA Diamond, a graduate-level science test, it scores 89.0 against 93.5 for Kimi K3. Meta also cites an Artificial Analysis Intelligence Index of 39, below GLM 5.3 (45) and Kimi K3 (44). So "most capable US open-weight model" is the claim, not "most capable open-weight model."

What the licence allows

Muse Spark Open ships under the Muse Model License, not an open-source licence such as Apache 2.0. Downloads are gated: Hugging Face asks for your legal name, date of birth and organisation. Products built on it must display "Built with Muse Spark Open." Fine-tuning and distillation are allowed, Meta claims no ownership of outputs, and using the model inside your own product or for research is free. The main commercial limit targets resellers: a company offering the model to others as a service must sign a separate agreement with Meta once that business passes $20 million in gross revenue over 12 months.

What's Still Unknown

Independent benchmarks. Every capability number so far is Meta's own. Artificial Analysis had not listed the model at launch, and Meta's full safety and evaluation report is described as "available soon."
Quality at smaller sizes. Community quantizations that squeeze the model onto a single Mac or a smaller GPU box will appear quickly. How much accuracy survives at 2 or 3 bits for an MoE model with 256 experts has not been measured.
Long-context speed on a desk. Apple's headline figures use an 8,192-token prompt. The model supports one million tokens, but prompt processing at 65,536 tokens already took about 40 seconds on four Macs, and nobody has published real numbers near the full window on Apple hardware.
Licence edge cases. Where "powering a feature in your own product" ends and "offering the model as a service" begins will matter to startups building on top of it, and Meta has not published guidance beyond the licence text.

Frequently Asked Questions

What is Muse Spark Open?

Muse Spark Open is an open-weight AI model released by Meta on October 15, 2026. It has 816 billion total parameters, about 42 billion active per token, a one-million-token context window, and accepts text, image, video and audio input. Meta built it for coding and agentic tasks and calls it the most capable open-weight model made in the United States.

Can Muse Spark Open run on a single Mac Studio?

Not at full 4-bit quality. Apple sizes the NVFP4 weights at about 493 GB, and Meta's model card recommends at least 700 GB of memory, while one M5 Ultra Mac Studio offers up to 512 GB. Meta and Apple's tested setup uses two Mac Studios linked by Thunderbolt 5. Smaller community quantizations may fit one machine, with unmeasured quality loss.

How fast is Muse Spark Open on Mac Studios?

Apple measured the 4-bit model with an 8,192-token prompt. Two M5 Ultra Mac Studios generated 15.74 tokens per second with pipeline parallelism and 24.99 with tensor parallelism. Four Mac Studios with tensor parallelism reached 36.73 tokens per second, with a 4.8-second wait for the first token. Longer prompts take noticeably longer to process before output begins.

Is Muse Spark Open free for commercial use?

Mostly. The Muse Model License lets companies use it in their own products, fine-tune it and keep their outputs at no cost, provided they display "Built with Muse Spark Open." Businesses that resell access to the model as a service must sign a separate agreement with Meta once that service passes $20 million in gross revenue over any 12-month period.

Is Muse Spark Open the same as Meta's Muse Spark?

No. Muse Spark is Meta's proprietary model, launched in April 2026, which powers Meta AI in WhatsApp, Instagram and Facebook and is not downloadable. Muse Spark Open is a separate model whose weights anyone can download after accepting the licence. Bloomberg described the release as Meta opening one model while keeping its best technology closed.

How does Muse Spark Open compare with Chinese open models?

On Meta's own figures it leads on tool-use tests such as MCP-Atlas (89.2) but trails on coding and reasoning. It scores 76.4 on Terminal-Bench 2.1 against 85.0 for Kimi K3 and 83.9 for GLM 5.3, and 39 on the Artificial Analysis Intelligence Index against 45 for GLM 5.3. Independent tests have not yet been published.

What is RDMA over Thunderbolt 5 and why does it matter here?

RDMA, remote direct memory access, lets one Mac write directly into another Mac's memory without going through the usual networking stack, cutting latency sharply. Apple added it in macOS 26.2. It is what makes tensor parallelism practical across Macs, because that method needs many tiny data exchanges for every layer of every token generated.

Where can developers get Muse Spark Open?

The weights are on Hugging Face in BF16 and NVFP4 formats after a gated sign-up, and through Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted APIs are offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. It runs in vLLM, SGLang, TensorRT-LLM, llama.cpp and Apple's MLX.