Muse Spark Open: High-Performance Agentic Model Released
The 816B parameter open-weight model enables advanced agentic workflows on desktop hardware.
1Launch of Muse Spark Open
Muse Spark Open is now available as a natively multimodal model featuring 816B total parameters and 42B active parameters. Designed specifically for coding and complex agentic tasks, it supports a 1M-token context window. This release marks a significant milestone in making high-capability models accessible for local execution on desktop hardware environments.
2Desktop Supercomputing Capability
The model is optimized to run on two Apple M5 Ultra Mac Studios using the MLX framework. By utilizing Apple’s unified memory architecture, the setup provides a combined terabyte of memory. The 1.2 TB/s bandwidth allows these desktop machines to function as a single supercomputer, handling the model's massive parameter requirements effectively.
3Agentic Workflow Performance
Testing demonstrated the model's ability to execute complex workflows entirely on-device. It successfully managed end-to-end tasks like shipping iOS updates for the Ice Cubes client and automating DaVinci Resolve video editing. These results highlight the model's capacity for multi-step reasoning and long-horizon software engineering tasks without needing cloud-based processing power.
4Broad Inference Engine Support
Muse Spark Open supports leading inference engines including SGLang, vLLM, and TensorRT-LLM. It features custom, auto-tunable Helion GPU kernels in SGLang, which provide significant speedups on NVIDIA and AMD hardware. Developers can also utilize PyTorch for further customization, ensuring compatibility across diverse silicon architectures like Intel, AMD, and Apple’s proprietary chips.
5Deployment and Access Options
The model weights are available for download on platforms like Hugging Face, Amazon SageMaker AI, and Google Cloud Model Garden. Additionally, developers can access the model through API platforms such as Groq, Together AI, and Fireworks AI. This multi-channel release strategy ensures that both local and cloud-based deployment options remain highly accessible.
6Safety and Evaluation Standards
Safety remains a priority, with the model undergoing post-training safeguards to address potential risks. Evaluations were conducted according to the Meta Superintelligence Scaling Framework, ensuring the model meets strict thresholds before public release. Detailed safety results are documented in the model card, with a comprehensive evaluation methodology report provided for transparency.
7Licensing and Customization
The weights are released under the Muse Model License, which includes limited commercial terms allowing free use for many developers. Users can customize the model for specific use cases by leveraging PyTorch’s TorchTitan. For Apple silicon users, the MLX framework facilitates direct weight integration, simplifying the local deployment process for advanced research.
8Strategic Impact of Release
Releasing Muse Spark Open is part of a broader effort to advance safety and bring personal superintelligence to the public. By enabling complex agentic work on desktop hardware, the release shifts the paradigm of model accessibility. This move encourages community innovation, inviting developers to share their creations on the Hugging Face forums.