Skip to main content

Kaer NewsSoftware

Meta releases Muse Spark Open, tuned by Apple for two Mac Studios

Meta released Muse Spark Open, an 816-billion-parameter open-weight model that Apple helped optimise to run on two M5 Ultra Mac Studios with 1TB of unified memory on 15 October 2026.

3 min read
Meta releases Muse Spark Open, tuned by Apple for two Mac StudiosPhoto: The Verge

Key points

Meta released Muse Spark Open, an 816-billion-parameter model optimised with Apple to run on two M5 Ultra Mac Studios with 1TB of memory.

Meta released Muse Spark Open on 15 October 2026, an open-weight AI model with 816 billion total parameters. Apple helped optimise it to run on two M5 Ultra Mac Studio machines, and Meta calls it the most capable US open-weight model to date and the first in its class that can run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.

The release matters because Meta and Apple, rivals that have argued for years over privacy and business interests, worked together on the launch and publicly backed each other. Meta chief AI officer Alexandr Wang said the two companies have been collaborating over several months. The model is aimed at long-horizon tasks, complex software engineering and multi-step reasoning, and it is downloadable now from Hugging Face.

What the release changes

Muse Spark Open uses a Mixture-of-Experts architecture with 256 experts, of which eight are activated per token, leaving about 42 billion parameters active for each token processed. Its context window holds one million tokens. The model takes text, images, video and audio as input but outputs text only. Meta promotes it for programming tasks, multi-step reasoning, and agents that use tools, correct errors and work on longer tasks.

The hardware requirement is the binding constraint. Apple's unified memory architecture lets the M5 Ultra's integrated GPU use up to 512GB per Mac Studio, so two machines can hold an entire model, Meta explains, making two Mac Studios behave like one supercomputer with a terabyte of unified memory. The 512GB Mac Studio is due in late October, with no specific price announced yet.

How two Mac Studios hold it

On benchmarks, Artificial Analysis gives Muse Spark Open 39 points on its Intelligence Index, the highest score among US open-weight models. Nvidia's Nemotron 3.5 Super V L, released at almost the same time, scores 26 with 120 billion parameters. Chinese models lead overall, and GLM-5.3, of similar size, reaches 45.

Where it ranks on benchmarks

Agent tasks are the weak point. Among the nine leading open-weight models, Muse Spark Open places last overall across four agent tests and fourth on the other six, and Artificial Analysis attributes its entire six-point gap to GLM-5.3 to those agent tests. Its strengths lie elsewhere: it scores 25 on the AA-Omniscience benchmark, top among open-weight models, with a 24% hallucination rate on questions it fails to answer correctly.

Meta's own benchmarks suggest the model outperforms two leading Chinese models, GLM 5.3 and Kimi K3, on agentic tasks and visual reasoning — a claim that sits beside Artificial Analysis's finding that it trails on agent tests. The model also uses fewer output tokens than any of the eight other leading open-weight models, averaging 40,000 per Intelligence Index task.

The Muse Model License permits commercial use but requires the notice "Built with Muse Spark Open" and compliance with a usage policy. Companies offering the model as a service through APIs need a separate agreement with Meta if they exceed $20 million in gross revenue over twelve months, counting revenue of affiliated companies.

Meta CEO Mark Zuckerberg wrote on X that the weights were releasing on 15 October 2026 and that the model runs well on new Mac Studios thanks to work with Apple CEO John Ternus. Ternus replied that Mac Studio keeps pushing the boundaries of on-device AI. The 512GB Mac Studio machines required for the two-machine setup are due in late October.

Frequently asked questions

What is Muse Spark Open?

Muse Spark Open is an open-weight AI model from Meta with 816 billion total parameters, released on 15 October 2026. It uses a Mixture-of-Experts design with 256 experts, activates about 42 billion parameters per token, and handles text, images, video and audio while outputting text only.

How much hardware do I need to run it?

You need two M5 Ultra Mac Studios with 512GB of unified memory each, giving 1TB total, linked over Thunderbolt 5 with Apple's MLX framework splitting the work. Alternatively, four 256GB Mac Studios work and cost roughly $38,000. The 512GB machine is due in late October.

How does Muse Spark Open compare with other models?

Artificial Analysis scores it 39 on its Intelligence Index, the top US open-weight score. It ranks last among nine leading open-weight models on agent tasks, while Meta says its own benchmarks show it beating GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning.

More from this edition