Meta releases Muse Spark Open, an 816B model that runs on two Mac Studios

Meta released the weights for Muse Spark Open on October 15. The model has 816 billion parameters, with 42 billion active for each token, and a 1-million-token context window. In its announcement, Meta says the model runs on two Apple M5 Ultra Mac Studios. One Mac Studio cannot hold the whole model.

Two 512GB Mac Studios are needed to hold the model

Each M5 Ultra Mac Studio can give its integrated GPU up to 512GB of unified memory. Meta’s post says the pair works like one machine with about a terabyte of memory, and that the chip’s 1.2 TB/s of memory bandwidth is fast enough to read the small slice of the model each token needs. The post does not say how the two machines are connected. Apple says Mac Studios can be clustered over Thunderbolt 5.

A smaller model needs far less memory. An Ollama setup runs on a Mac with 8GB of RAM.

Meta says it ran agentic workflows entirely on-device with Muse Code, its coding agent, powered by Muse Spark Open. The tests included iOS fixes and new features for Ice Cubes, an open-source Mastodon client. In a DaVinci Resolve test, the model reviewed footage frame by frame and cropped clips to vertical format.

The 512GB configuration is due in late October

Apple lists $5,499 as the starting price of the M5 Ultra Mac Studio. The 512GB configuration is coming in late October. Meta’s model card puts the NVFP4 version at 700GB of memory, so the two-machine setup likely needs that top option.

The Information reported that demand for local AI hardware pushed Apple’s August refresh forward. Apple has not confirmed that.

A comparison of the M6 Mac mini and M5 Ultra Mac Studio finds the Mac mini the better choice for most people. The Studio suits heavier work such as large-scale AI.

Hugging Face needs a login, and large API resellers need a Meta deal

The weights are on Hugging Face in BF16 and NVFP4 versions. The page asks users to log in and review the conditions before they can access the model. Meta’s model card says running the model on your own system requires at least 700GB of GPU memory for NVFP4 and 2TB for BF16. The card does not say which version the Mac setup uses.

The model ships under the Muse Model License. Meta says its commercial terms are limited and let many developers use the model for free. The license sets a revenue threshold for hosted services. A company that gives third parties API access to the model must sign a separate written agreement with Meta once its combined gross revenue passes $20 million in any 12-month period. Internal use and research are exempt.

Readers without two Mac Studios can use hosted API providers, including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models also offer the weights for download.

Meta’s own numbers put it top on MCP-Atlas but behind some rivals overall

Meta’s model card compares Muse Spark Open with other open-weight models. It scores 39 on the Artificial Analysis Intelligence Index, behind Qwen3.8 at 40, Kimi K3 at 44, and GLM 5.3 at 45. It scores 89.2 on MCP-Atlas, a test of agent tool use, the highest score in that table.

Because the weights are open, Meta added extra safety training, and the model met the release thresholds in its Meta Superintelligence Scaling Framework. A full evaluation report is “available soon.”

Meta said in April it hoped to open future Muse models

Meta launched Muse Spark in April to power the Meta AI app and meta.ai. The April post said Meta hoped to open-source future versions of the model. The Muse Spark Open model card describes the release as an open-weights version with additional safety post-training. Meta’s announcement does not give a date for the next open release.