Meta on Thursday released Muse Spark Open, an 816-billion-parameter open-weight model that the company calls the most capable US model of its kind. Announcing the release on X, Meta said the model can run at home on a pair of Apple M5 Ultra Mac Studios.

Meta's model card states that Muse Spark Open uses a mixture-of-experts design that activates 42 billion of its 816 billion parameters per token — a sparse setup that keeps compute per token tied to the active slice, even though the hardware still has to hold the full weight set. The model accepts text, images, video and audio as input and returns text, with a context window of up to 1 million tokens. Meta aims the model at coding and agentic work, targeting long-horizon tasks, software engineering and reasoning across multiple steps.

Two Mac Studios, about a terabyte of memory

Each M5 Ultra Mac Studio carries up to 512GB of unified memory, so two machines together offer roughly a terabyte — which Meta said in its announcement was enough to store the entire model. The model runs locally on Apple's MLX framework. Memory is the deciding constraint for hardware like this: the model's stated requirements sit far beyond what any single GPU offers, which is why the reference setup spans two machines rather than one accelerator.

Apple will ship the 512GB configuration in late October, but it has not said how much it will cost — a figure that will set the real price of the two-box rig Meta described. The 256GB M5 Ultra Mac Studio currently costs $10,799, meaning four units — about a terabyte in total — would cost $43,196 before tax. Meta's model card says the NVFP4 build of Spark Open requires at least 700GB of GPU memory, while the BF16 build needs at least 2TB.

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.

Spark scores 39 on Artificial Analysis, below three Chinese rivals

Meta's "most capable" claim covers only US models. The company's own table gives Spark Open a score of 39 on the Artificial Analysis Intelligence Index, where several open-weight Chinese models rank higher: Xiaomi's MiMo-V2.6-Pro sits at 46, GLM-5.3 at 45 and Moonshot AI's Kimi K3 at 44.

Spark Open leads its listed rivals on JobBench at 61.8 and MCP-Atlas at 89.2, both tool-use tests, as well as 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. By Meta's own table, the release tops the US open-weight field yet stops short of the overall open-weight lead — a gap the broader usage numbers put into perspective.

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," Interconnects author Nathan Lambert wrote on X. He added, "Still more work to do!"

Hosting and licensing

Seven partners host Spark Open: Together AI, Fireworks AI, Baseten, Groq, Modal, Nebius and OpenRouter, giving developers API access without downloading the weights themselves. The weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta offers the model under the Muse Model License, which permits many developers to use it for free but with limited commercial terms.

The release reverses Meta's April approach, when Muse Spark launched as 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.