Meta Releases Muse Spark Open, An 816-Billion-Parameter Open-Weight AI Model For Mac Studios

Meta has unveiled Muse Spark Open, an 816-billion-parameter open-weight AI model tailored for coding and agentic tasks, claiming it to be the most powerful open-weight model from the United States. Introduced on October 15 via Hugging Face, top cloud service providers, and multiple API providers, the model boasts a context window of one million tokens and operates natively in a multimodal format.

Designed for coding and complex software engineering, it can handle long-running tasks and multi-step reasoning. Meta enhanced the model with additional safety training to minimize misuse when deployed without guardrails. The company collaborated with Apple to optimize running the model locally on the M5 Ultra Mac Studio through MLX, Apple's machine-learning framework for its own silicon.

Users can connect two to four Mac Studios to harness the model's capabilities for coding or agentic work. Developers can access the model through API platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

The model can be utilized with various inference engines, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth, and MLX, which Meta has optimized for AMD, Apple, Intel, and NVIDIA hardware.