⏱️ Reading time: 8 min
Meta just released the weights for Muse Spark Open, an 816 billion parameter model that runs entirely on two Mac Studio units with the M5 Ultra chip, without going through any cloud server.
📑 En este artículo
It’s the open-weight version of the system that Meta Superintelligence Labs introduced a few weeks ago, and it arrives with a 1 million token context window designed for multi-step software engineering and extended agentic work.
TL;DR
- Meta published the weights for Muse Spark Open, an MoE model with 816 billion total parameters and 42 billion active.
- The context window reaches 1 million tokens, designed for long software engineering tasks.
- Two Mac Studio units with the M5 Ultra chip and MLX run the full model thanks to 512 GB of unified memory per machine.
- Helion kernels in SGLang speed up inference by 1.03 to 3.95 times on Nvidia and AMD GPUs.
- The model is available on Hugging Face and via APIs from Together AI, Fireworks, Baseten, Groq, Modal, Nebius, and OpenRouter.
What Muse Spark Open Is
Muse Spark Open is the open-weight model derived from Muse Spark, Meta’s multimodal mixture-of-experts system with 816 billion total parameters, 42 billion active per token, and a 1 million token context window, designed for coding and multi-step agentic work.
The model is natively multimodal: it processes text and images within the same mixture-of-experts (MoE) architecture, and activates only 42 billion of the 816 billion parameters per token. That partial activation is what sets an MoE apart from a dense model, since the compute cost per token resembles that of a much smaller model, even though the memory needed to store all the experts is still that of an 816 billion parameter model.
What Happened
Meta’s open-weight model is now available for download on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. For those who’d rather not host the model, seven platforms serve it as an API: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
Muse Spark Open runs on the most widely used inference engines right now: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with specific optimizations for AMD, Apple, Intel, and Nvidia silicon. Anyone looking to fine-tune the model for their own use case can do so with PyTorch, using the TorchTitan distributed training framework.
Context and History
Meta Superintelligence Labs introduced the closed version of Muse Spark a few weeks ago as the lab’s first model. The news in October isn’t the model itself, but that Meta decided to open its weights, presenting it, according to the announcement itself, as the most capable open-weight model built in the United States to date.
The move places Meta’s open-weight model in a race where China had been setting the pace: open models like those from Moonshot AI, Z.ai, or DeepSeek had spent months leading open-weight rankings on benchmarks like Artificial Analysis. Meta responds with a large model, but with a different pitch: it runs on desktop hardware, something no rival of that size had offered until now.
Technical Details: How the Two Mac Studio Units Split the Memory
The technical reason why two Mac Studio units are enough to run an 816 billion parameter model lies in Apple’s unified memory architecture. Each M5 Ultra chip lets the integrated GPU use up to 512 GB of that memory, so between the two machines they add up to a terabyte available for storing the full model’s weights.
Bandwidth is the other half of the equation. The M5 Ultra delivers 1.2 TB/s of memory bandwidth, a figure that’s enough because an MoE model doesn’t need to read all 816 billion parameters at every step: with the router choosing which experts to activate, each token only needs to read the 42 billion parameter slice assigned to it. MLX handles splitting that slice between the two machines so that, from the model’s perspective, they behave as a single machine with double the memory.
flowchart TD
A["Muse Code Agent"] --> B["Model's MoE Router"]
B --> C["Mac Studio 1: M5 Ultra, 512 GB"]
B --> D["Mac Studio 2: M5 Ultra, 512 GB"]
subgraph "Combined unified memory: ~1 TB"
C
D
endOn that basis, Meta ran full agentic tests on local hardware, without touching the cloud: using Muse Code, its coding assistant, on top of the open-weight model, the team applied real changes to Ice Cubes, an open source Mastodon client for iOS, including bug fixes, new features, and a full build. In another test, the model handled DaVinci Resolve end to end, reviewing footage frame by frame to find the best moments and cut them into vertical format.
On the other end of the hardware spectrum, for those running the model on data center GPUs, Meta published its own GPU kernels called Helion within SGLang. Those self-tuning kernels deliver between 1.03 and 3.95 times more speed at the kernel level on Nvidia and AMD GPUs, and a 3% end-to-end improvement on an AMD MI350.
| Option | When to Use It | Advantage | Limitation |
|---|---|---|---|
| Cloud API (Together AI, Fireworks, Groq, etc.) | To integrate the model without managing hardware | No upfront investment, scales on its own | Depends on the provider’s availability and pricing |
| Own weights on data center GPU | For high volume or fine-tuning with TorchTitan | Full control, and Helion kernels speed up inference | Requires high-end Nvidia or AMD GPUs |
| Two Mac Studio M5 Ultra units locally | To run the full model without cloud or server GPUs | Total privacy, no recurring API cost | The hardware costs several thousand dollars and needs maintenance |
💡 Tip: before buying the hardware, try Meta’s open-weight model via API (Together AI or OpenRouter, for example) to confirm the use case justifies it.
Impact and Analysis
For a development team, the practical difference isn’t the model’s size but where it runs. An 816 billion parameter model served via API isn’t news in 2026: what’s different is that there’s now an option to run it in an office, without a project’s source code or footage ever leaving the local network. That matters for teams with confidentiality commitments or that simply can’t send sensitive data to an outside provider.
Licensing is also part of the calculation. Meta releases the model under the Muse Model License, with limited commercial terms that let most developers use it for free, though it isn’t an open source license in the strict sense: companies above a certain user threshold tend to fall outside the free tier, the same pattern Meta already used with the Llama family.
📌 Note: open weights isn’t the same as open source. Meta’s license restricts commercial use past a certain scale, something worth checking before integrating Meta’s open-weight model into a product.
The other angle is safety. Meta says it evaluated the model under its Meta Superintelligence Scaling Framework and applied additional safety training given that it’s open weight, since once weights are published there’s no way to revoke them or patch the model remotely.
What’s Next
Meta promised a full evaluation report soon, in addition to the model card already published on Hugging Face. The company invites developers to share work built with the model on Hugging Face forums, the channel it typically uses to gauge real adoption before iterating on the next version.
If adoption grows, the next logical chapter is the same one Llama and the Chinese open models already went through: community fine-tunes for specific tasks, and performance comparisons against Kimi K3, GLM-5.2, or Qwen3.8 on independent benchmarks.
If you have two Mac Studio M5 Ultra units on hand, the shortest path to trying it is installing the MLX adapter and downloading the weights directly from the Muse Spark Open collection on Hugging Face.
Frequently Asked Questions
What is Muse Spark Open and who released it?
It’s the open-weight model that Meta Superintelligence Labs derived from its Muse Spark system, with 816 billion total parameters and 42 billion active per token.
What hardware do I need to run Meta’s open-weight model at home?
According to Meta’s announcement, the validated setup is two Mac Studio units with the M5 Ultra chip, using the MLX framework to split the model between both machines.
How does Meta’s open model differ from the closed version of Muse Spark?
They share architecture and size, but the open version publishes the weights for download and self-hosting, while the closed version is only used through Meta’s own products.
What license does Meta’s open MoE have?
It’s distributed under the Muse Model License, with limited commercial terms that allow free use for most developers.
Where can you try Meta Superintelligence Labs’ open AI?
Via API on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, or by downloading the weights from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, or Microsoft Foundry Models.
References
- Meta AI Research: official announcement of the model, with benchmarks and licensing details.
- Hugging Face: collection with the model’s downloadable weights.
- GitHub: ml-explore/mlx: official repository of the MLX framework used to run the model on Apple Silicon.
- PyTorch: framework used to customize the model with TorchTitan.
📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion
Featured image: Foto de Google DeepMind en Unsplash
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments