Meta has released Muse Spark Open, making the model’s weights available through Hugging Face for developers to run and customize, with optimization for Mac Studio systems using M5 Ultra. Meta says two of those machines can hold the entire model, allowing it to run on desktop hardware. Access is available now under the Muse Model License, whose limited commercial terms permit free use by many developers.
Developers can run the model on Apple silicon through MLX, which retrieves its weights from Hugging Face. For developers who want hosted inference, Meta lists providers including OpenRouter and Together AI. Developers can also adapt the model to their needs through further tuning with PyTorch’s TorchTitan.
Read more: Meta Releases Muse Wonder With Memory Across Conversations
The Mac configuration draws on Apple’s unified memory design, which gives the integrated GPU access to as much as 512 GB in each Mac Studio. Meta puts the M5 Ultra’s memory bandwidth at 1.2 TB/s and says a pair of systems provides a terabyte of unified memory. The company says it used the latest MLX version to optimize Muse Spark Open for that hardware.
Meta also ships Helion GPU kernels that can tune themselves within SGLang, with reported kernel speed gains ranging from 1.03 to 3.95 times on GPUs from AMD and NVIDIA. Its reported gain for the complete workload on AMD MI350 is 3%, a separate measure from those kernel results.
In local testing, Meta used Muse Code with Muse Spark Open to complete iOS development work on Ice Cubes, an open-source Mastodon client, including fixes, features and a complete build. Another test involved operating DaVinci Resolve, examining individual video frames to select highlights and produce vertical crops.
The release follows Muse Glimmer, a 30-billion-parameter model that Meta released in August for use on a single consumer GPU under Apache 2.0. In September, Meta released its proprietary Muse Spark 1.3 with improvements aimed at coding and agentic work.
Meta says Muse Spark Open received additional safety post-training because its weights would be publicly available. The company says evaluations under its Meta Superintelligence Scaling Framework found that the model satisfied its release thresholds for the most severe risks. Meta has provided detailed findings in the model card and says a fuller evaluation report will follow soon.
