Meta has released Muse Spark Open, an open-weight model with 816 billion total parameters and 42 billion active parameters per token, designed for coding and agentic work with a 1 million-token context window Meta AI Research.
The release extends Meta's lineup of open-weight models into a larger, multimodal category, giving developers a model that accepts text, images, documents, audio, and audio-video inputs and outputs text. The model uses a mixture-of-experts (MoE) architecture and is available in two checkpoint formats, BF16 and NVFP4, to suit different hardware Meta Model API.
Weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models, and the model can be deployed with inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth Meta AI Research. Meta said it is optimized for AMD, Apple, Intel, and NVIDIA silicon Meta Model API.
Meta said Muse Spark Open has been optimized for Apple Silicon, using the latest MLX version to run on Mac Studio with M5 Ultra, and was tested on agentic workflows including real iOS changes and video editing tasks Meta AI Research. The model also includes custom, auto-tunable Helion GPU kernels in SGLang, which Meta said deliver significant speedups on NVIDIA and AMD GPUs Meta AI Research.
Meta said the model underwent additional safety post-training and was evaluated using its Superintelligence Scaling Framework Meta AI Research. It is released under the Muse Model License with limited commercial terms, allowing many developers to use it for free Meta AI Research.
The announcement does not include independent benchmark results or comparisons with other open-weight models, and no peer-reviewed evaluation is cited.