Meta Releases Muse Spark Open, a New Open-Weight AI Model for Coding

Meta Releases Muse Spark Open, a New Open-Weight AI Model for Coding

Meta announced a new open-weight artificial intelligence model on Thursday. The model, called Muse Spark Open, targets coding, software development, and complex multi-step tasks. It has hundreds of billions of parameters and a large context window.

Meta described Muse Spark Open as the most capable US open-weight model to date. The model has 816 billion total parameters, with 42 billion active parameters. It supports a one-million-token context window. Meta designed it for coding and agentic workloads, including long-running tasks, complex software engineering, and multi-step reasoning.

Meta partnered with Apple to bring the model to Mac Studio desktops powered by the M5 Ultra chip. Developers can run agentic AI workflows locally on these machines. Meta optimized the model using Apple’s MLX machine learning framework. This setup reduces the need for full dependence on cloud-based inference.

The system uses Apple’s unified memory architecture. Each M5 Ultra Mac Studio provides up to 512GB of memory. Two systems together offer up to 1TB of unified memory. This allows the machines to run the full model weights locally.

Muse Spark Open supports several inference engines. These include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, and llama.cpp. The model can run on hardware from AMD, Apple, Intel, and Nvidia. Developers can also customize it with PyTorch and TorchTitan.

The model includes custom, auto-tunable Helion GPU kernels in SGLang. Meta reported up to 3.95 times kernel-level speed improvements on Nvidia and AMD GPUs. It also reported a 3 percent end-to-end speed improvement on AMD MI350 hardware.

Developers can access Muse Spark Open through API platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. They can also download the weights from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

Meta released the model under its Muse Model License. The license allows free use by developers under certain commercial terms. Meta said it conducted safety evaluations under its Superintelligence Scaling Framework and that the model meets the required release thresholds.