Meta releases Muse Spark Open weights

Meta says Muse Spark Open is natively multimodal and built for coding and agentic work, including long-horizon tasks, complex software engineering and multi-step reasoning.[1] The model has 816 billion total parameters and 42 billion active parameters, according to Meta.[1] Meta also lists a 1-million-token context window.[1] These are the announcement’s headline specifications; teams evaluating the model can use them to identify workloads for testing, rather than assume performance on their own tasks.

Meta says the model brings agentic capabilities to desktop computers and is the first in its class to run on a desktop using two Apple M5 Ultra Mac Studios and MLX.[1] Meta says each Studio can provide up to 512 GB of memory for the model, and that two can hold the entire model.[1] The company also says Muse Spark Open drove changes to the Ice Cubes iOS app and operated DaVinci Resolve end to end, reviewing footage frame by frame.[1] These examples describe Meta’s demonstrations, not a guarantee of results in other workflows.

Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as API platforms offering Muse Spark Open.[1] Its weights are available for download through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.[1] Meta says the model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth.[1] Its inference-engine optimizations support AMD, Apple, Intel and NVIDIA silicon, according to the company.[1]

Meta says the weights are released under the Muse Model License, which has limited commercial terms, while allowing many developers to use the model for free.[1] It also says the model received additional safety post-training and met the thresholds its framework requires for release.[1] Meta reports kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD GPUs, and says Helion GPU-kernel work produces a 3% end-to-end speedup on AMD MI350.[1] Those figures are Meta’s reported results; they do not establish performance for every configuration.