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

Muse Spark Open is the biggest open-weight release of the autumn so far. On October 15, Meta Superintelligence Labs published the weights of Muse Spark, the flagship model that until now was only reachable through Meta's own API, and it showed the full model running on a pair of Apple Mac Studios. For researchers, startups and anyone who wants frontier-class AI on hardware they control, this is a genuinely new option.

• Model size: a mixture-of-experts design with about 816B total parameters and about 42B active per token, plus a 1M-token context window, per Meta.
• Inputs: natively multimodal, taking text, images, audio and video in and producing text out, according to the Hugging Face launch post.
• Desktop path: Meta says it runs on two Apple M5 Ultra Mac Studios (512GB each, roughly a terabyte combined) using MLX.
• License: the Muse Model License is free for many uses but requires "Built with Muse Spark Open" attribution and a separate deal for large hosting providers.

What Is Muse Spark Open?

Muse Spark has been Meta's top-tier model since spring, and we have followed it through several upgrades, most recently Muse Spark 1.3's frontier benchmark results. Muse Spark Open is that same model family released as downloadable weights, with extra safety post-training. Meta says it met the release thresholds of its internal scaling framework, and a full evaluation report is promised soon.

Under the hood it is a sparse mixture-of-experts model. The Hugging Face post explains that each layer holds 256 experts and selects 8 per token, which is why only about 42B of the 816B parameters do work on any single step. That keeps generation fast, though every expert still has to sit in memory.

How Well Does Muse Spark Open Perform?

Meta's model card lists strong agentic and reasoning numbers at its highest reasoning setting:

• Terminal-Bench 2.1: 76.4
• GPQA Diamond: 89.0
• OSWorld Verified: 80.4
• MCP-Atlas: 89.2
• Humanity's Last Exam with tools: 61.0
• Artificial Analysis Intelligence Index: 39

These are Meta's published results, and Meta describes the release as the most capable open-weight model from a US lab. Independent leaderboards will tell the fuller story over the coming weeks, but the tool-use scores in particular point to a model built for coding agents and computer-use workflows, which is where the open-weight world has wanted a stronger option.

Can You Really Run Muse Spark Open on a Mac?

Yes, with the right Macs. Meta's research post says the model runs across two M5 Ultra Mac Studios with the latest MLX framework, each machine contributing up to 512GB of unified memory at 1.2 TB/s. TechCrunch reports Meta demonstrated it with Muse Code, its AI coding terminal, fixing bugs and adding features to a Mastodon client, and reviewing footage in DaVinci Resolve.

If you are working with a single large Mac, quantized builds help. Hugging Face reports a 2-bit llama.cpp quant used about 304GB and ran at 18 tokens per second on a 512GB M3 Ultra, while a 3-bit MLX quant takes about 344GB. Our guide to running local LLMs on a Mac with MLX covers the setup basics.

Where Can Developers Get It Today?

Availability on day one is unusually broad:

• Weights: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
• Hosted APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
• Data-center hardware: the model card cites at least 700GB of VRAM for NVFP4 and 2TB for BF16; Hugging Face lists a single B200 node for NVFP4 in Transformers.

The license deserves a careful read. Meta claims no ownership of outputs, internal and research use is open, and the extra agreement only applies to model-as-a-service providers with more than $20 million in annual revenue.

Why This Release Matters

Open weights at this scale change who gets to experiment. University labs can study a frontier-class model directly, companies with strict data rules can deploy it on their own infrastructure, and the local AI community gets a new ceiling to aim for. It is a strong moment for open model development, and you can follow what comes next in our AI news coverage.

Sources: Meta AI Research: Muse Spark Open — October 15, 2026; Hugging Face: Welcome Muse Spark Open — October 15, 2026; Hugging Face model card — October 15, 2026; TechCrunch — October 15, 2026.