Meta on Thursday pushed out Muse Spark Open, an 816-billion-parameter open-weight model that it says is the most capable US model of its kind.
The model runs on a pair of Apple M5 Ultra Mac Studios at home, the company says.
Each M5 Ultra brings up to 512GB of unified memory
Meta’s model card says that Muse Spark Open uses a mixture-of-experts design that activates 42 billion of its 816 billion parameters per token.
It accepts text, images, video and audio and returns text. It has a context of up to 1 million tokens.
Meta aims it at work that involves code and agents. Its goals are long-horizon tasks, software engineering and reasoning in more than one step.
Each M5 Ultra Mac Studio has up to 512GB of unified memory, so two would have about a terabyte, which Meta said in its announcement was enough to store the entire model. It runs locally on Apple’s MLX framework.
In Meta’s demos, Muse Code, built on Spark Open, shipped changes to Ice Cubes, an open-source Mastodon client. It also drove the DaVinci Resolve video editor to scrutinize footage and cut vertical crops.
Apple will ship the 512GB model in late October, but it hasn’t said how much it will cost yet. Right now, a 256GB M5 Ultra Mac Studio costs $10,799. That means four of them, which is about a terabyte, would cost $43,196 before tax.
Meta’s model card says the NVFP4 build needs at least 700GB of GPU memory and BF16 needs at least 2TB.
Spark scores 39 on Artificial Analysis, below three Chinese rivals
Meta’s “most capable” claim only covers US models. Meta’s own table gives Spark Open a score of 39 on the Artificial Analysis Intelligence Index.
Several open-weight Chinese models rank higher on the same list. While GLM-5.3 is at 45 and Moonshot AI’s Kimi K3 is at 44, Xiaomi’s MiMo-V2.6-Pro is at 46.
Spark Open leads its listed rivals on JobBench at 61.8 and MCP-Atlas at 89.2, both tool-use tests, and on HLE-Diamond without tools. Kimi K3 and GLM 5.3 outdo it on terminal coding, longer software-engineering tasks and the GDPVal-AA professional-work benchmark.
Chinese models accounted for 41% of downloads in the year to March 2026, outpacing US models, according to an assessment by Hugging Face. In September, OpenRouter said open-weight models accounted for 60% of the token volume from US-based requests, the majority from Chinese labs.
“I waited so long for US open weight models to be relevant again,” wrote Interconnects author Nathan Lambert on X. He added, “Still more work to do!”
Spark Open is hosted by seven partners: Together AI, Fireworks AI, Baseten, Groq, Modal, Nebius and OpenRouter. Weights are hosted on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
It’s offered by Meta under the Muse Model License, which permits many developers to use it for free, but with limited commercial terms.
The release overturns Meta’s April approach. Then Muse Spark launched closed and proprietary, under new AI chief Alexandr Wang, in what Meta called a full overhaul of its AI work.
Open weights put the parameters in anyone’s hands to download, run and fine-tune, the stance Meta held before the Muse Spark line.
The smartest crypto minds already read our newsletter. Want in? Join them.