Models
Meta's new open AI fits on two Macs but trails Chinese models
Meta has released Muse Spark Open, a model that anyone can download and that runs on two Apple Mac Studio computers. An independent measurement ranks it as the best open model from the US, but behind two Chinese models.

The essentials
- Meta released Muse Spark Open on October 15, 2026, with open weights: anyone can download it and run it on their own machines (Hugging Face Blog, TechCrunch).
- Claimed by one partyAccording to Meta, it runs on two Mac Studio computers with the M5 Ultra chip thanks to joint work with Apple, which has publicly praised it (The Verge, MarkTechPost).
- Artificial Analysis gives it 39 points: the best open model from the United States, but behind the Chinese models GLM-5.3 and Kimi K3 (MarkTechPost, Latent Space).
- On services such as OpenRouter or Together, it costs $1.25 per million tokens sent and $4.25 per million generated (MarkTechPost, Latent Space).
- The license requires users to display "Built with Muse Spark Open" and demands a separate agreement from anyone who resells it with revenue above $20 million (MarkTechPost, Latent Space).
Why it matters
An open model at this level fitting on a desktop changes where it can be used. That desktop is still very high-end. Even so, companies and programmers could work with it without sending their code to someone else's servers. The partnership with Apple, which usually competes with Meta, suggests that Apple wants to sell its Macs as machines for AI.
The independent measurement tempers the announcement. Meta trained it to work as an agent, and according to Latent Space it comes last among the nine best open models in Artificial Analysis's agent tests (Latent Space). If this is confirmed, the United States is closing the gap with China in open AI, but has not caught up yet.
If you don't write code, not much for now. Running it at home requires two top-of-the-line Mac Studio computers; the realistic option is to pay as you go on the services that already offer it.
The details
It is a huge model, with about 816 billion parameters, which are the internal values it adjusts as it learns (Hugging Face Blog). It doesn't use them all at once: for each piece of text, it activates about 42 billion (Hugging Face Blog). That makes it faster, but it means the whole model has to be loaded in memory (Hugging Face Blog). It understands text, images, audio and video, and responds with text (Hugging Face Blog). Meta trained it mainly to write code as an agent that works on its own on long tasks (Hugging Face Blog, NVIDIA Technical Blog).
The compressed version that Meta distributes takes up about 530 GB (MarkTechPost, Latent Space). According to Meta, it runs on two Mac Studio computers with the M5 Ultra chip and 512 GB of memory each, connected by Thunderbolt 5 (The Verge, MarkTechPost). Engadget, cited by TechCrunch, estimates that it needs at least one terabyte of memory, which matches those two machines (TechCrunch). Unsloth has released even more compressed versions that go down to about 300 GB (MarkTechPost).
Artificial Analysis, a company that compares models, gives it 39 points on its overall index: first among open models from the United States (MarkTechPost, Latent Space). But it ranks behind the Chinese models GLM-5.3, with 45, and Kimi K3, with 44 (MarkTechPost, Latent Space). In Meta's own tables, it stands out at using tools and falls behind at coding (MarkTechPost). According to Latent Space, it also performs worse than Spark 1.3, another of Meta's own models (Latent Space).
It can already be used on services such as Together, Fireworks, Baseten or OpenRouter (MarkTechPost, Latent Space). It costs $1.25 for every million tokens (pieces of text) you send it and $4.25 for every million it generates (MarkTechPost, Latent Space). The license allows people to use, modify and redistribute it, but it is not fully free: it imposes branding and revenue conditions (MarkTechPost, Latent Space). Nathan Lambert, a well-known voice in the industry, has sharply criticized it (Latent Space).
What we don't know
- How fast it actually responds on two Mac Studio computers.
- Whether the extra safety training that Meta mentions holds up when others modify the model.
- How much worse it performs than Spark 1.3, the other Meta model it is compared with.
Background
Muse Spark Open is the big brother of Muse Glimmer, a small model that Meta released in August (Hugging Face Blog). On October 15, we reported that a British NGO had tested 134 AI models with questions a terrorist might ask, and that versions of open models modified to remove their safeguards were appearing faster and faster. According to MarkTechPost, Meta says this open version received extra safety training to make it more resistant even without external controls (MarkTechPost).
How this story has changed
- We added the independent measurement by Artificial Analysis, which places it behind two Chinese models, the usage price, the exact memory it needs and the terms of its license.
- We published the story.
Sources
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
- Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
- Meta and Apple team up to make Muse AI models run better on Macs
- Meta launches an open-weight AI model that runs on Macs
- Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context
- [AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment
- Meta shakes up the frontiers, open and closed
Was this story useful? Yes Not really