Meta Releases Open-Weight Muse Spark Open: 816B Parameters, Runs on Two Mac Studios, Still Behind GLM 5.3 and Kimi K3 in Meta's Own Table

On October 15 Meta released Muse Spark Open, an open-weight model with 816 billion total and 42 billion active parameters, native multimodality and a 1-million-token context window. Weights are on Hugging Face, and the model is available by API through Baseten, Fireworks AI, Groq, OpenRouter, Together AI and others. Meta calls it the most capable US open-weight model to date and demonstrated it running locally on two M5 Ultra Mac Studios (about 1TB of unified memory combined) using MLX. It ships under the Muse Model License: users must display "Built with Muse Spark Open," and companies with more than $20 million in annual revenue must notify Meta and sign a separate agreement before offering it as a model-as-a-service. The comparison table in Meta's model card gives it 39 on the Artificial Analysis Intelligence Index, below GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40, though it leads on agent-style tests such as MCP-Atlas and JobBench.

What shipped

On October 15 Meta released Muse Spark Open, an open-weight model. It comes three days after Muse Wonder, the closed frontier model this site covered, and is the first time the Muse family has released weights. Specs: 816 billion total parameters with 42 billion active per token (a mixture-of-experts design), native multimodality and a 1-million-token context window, aimed at coding, agents and long-horizon tasks.

Weights are on Hugging Face and can also be downloaded from Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry; API access comes from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

What "runs on a Mac" actually takes

Meta's demo setup is two M5 Ultra Mac Studios linked with Apple's MLX framework, each with up to 512GB of unified memory, about 1TB combined. The model card lists memory needs of at least 700GB at NVFP4 precision and 2TB at BF16. In other words, "local" means a pair of top-spec, high-memory workstations, not an ordinary laptop or a single desktop.

What the license says

Open weights are not the same as an open-source license. The Muse Model License requires users to display "Built with Muse Spark Open" in their product UI or documentation. Companies with more than $20 million in annual revenue must notify Meta and sign a separate written agreement before offering the model as an external model-as-a-service; internal use, research and embedding it in their own products are not covered by that clause.

The limits of "most capable in the US"

Meta calls it the most capable US open-weight model to date, and the "US" qualifier matters. In the comparison table in Meta's own model card, Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, against 45 for GLM 5.3, 44 for Kimi K3 and 40 for Qwen3.8. Where it leads is tool-use and agent tests such as MCP-Atlas (89.2) and JobBench (61.8); on Terminal-Bench 2.1 it trails all three.

For anyone choosing a model, it adds an option with a clear license and full ecosystem support (SGLang, vLLM, llama.cpp and MLX are all supported). On overall scores alone, though, it is not at the top of the open-weight field.

via: Meta official blog, Hugging Face model card, TechCrunch report