Meta released Muse Spark Open on October 15, the open-weight version of Muse Spark, the model family behind its Meta AI assistant. Meta calls it "the most capable US open-weight model to date" and says it is built for coding and agentic work: long-horizon tasks, complex software engineering and multi-step reasoning.

Key numbers

Size: a mixture-of-experts model with 816 billion parameters in total, 42 billion of them active.
Context: 1 million tokens.
Input and output: it accepts text, images, video and audio and answers in text. The model card says it was trained on data in more than 100 languages.
Weights: BF16 and NVFP4 versions on Hugging Face, also offered through Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry.

Running it yourself

Meta says Muse Spark Open is the first model in its class that runs on a desktop setup of two Apple M5 Ultra Mac Studios with Apple's MLX framework; each Mac Studio can have up to 512 GB of unified memory. On GPUs, the model card asks for at least 700 GB of VRAM for NVFP4 and about 2 TB for BF16. Supported engines include vLLM, SGLang, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth, and fine-tuning works through PyTorch's TorchTitan.

Without that hardware, you can use the model through hosted APIs from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

The license

The weights come under the Muse Model License, which Meta describes as having limited commercial terms. Companies that offer the model itself as a service need to notify Meta and sign a separate agreement once their group's gross revenue exceeds $20 million over 12 months. That threshold does not apply to research, internal use, or using the model inside your own product that is not a model service. Products must display "Built with Muse Spark Open", use must follow Meta's usage policy, and the model is not intended for people under 18.

Meta's announcement gives no benchmark figures, so independent tests will have to show how it compares with closed models.

What this means if you're choosing a model

If you just want to chat with a Muse model, nothing changes: the Meta AI assistant in WhatsApp, Instagram and Facebook is still the easiest way, with no setup.

If your data has to stay on your own machines, Muse Spark Open is a model of this size you can run locally, though two top-end Mac Studios or a large GPU server is a serious investment. Read the license before you sell access to it.

If you want to test it cheaply, start with a hosted API such as OpenRouter or Groq on a real task from your work, and compare it with another open model family such as DeepSeek before you commit.