Meta releases Muse Spark Open, saying it is the most capable US open-weight model, with 816B total and 42B active parameters and a 1M-token context window

Your coding assistant capabilities improve significantly for long-horizon tasks, with no immediate change required from you.

Meta has released Muse Spark Open, an open-weight large language model featuring 816 billion total parameters and 42 billion active parameters. The model boasts a 1-million-token context window and is designed for coding and agentic tasks requiring long-horizon reasoning and complex software engineering. Notably, Muse Spark Open is optimized to run on Apple's M5 Ultra Mac Studios using the MLX framework, allowing it to function on desktop hardware. Meta is making the model weights available through various platforms including Amazon SageMaker, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. It also supports multiple inference engines and hardware accelerators, including NVIDIA, AMD, and Intel silicon. The release aims to democratize access to advanced AI models, enabling users to run them locally or through API partners like Baseten and Fireworks AI.

The release of Muse Spark Open signifies a major step in bringing highly capable, natively multimodal LLMs to consumer hardware, particularly Apple Silicon. By optimizing for the M5 Ultra Mac Studio and leveraging Apple's unified memory, Meta demonstrates that desktop environments can now handle models with massive parameter counts and context windows, challenging the notion that such power is solely cloud-bound. This move, coupled with its availability through various cloud providers and inference engines, broadens accessibility for developers and potentially enterprises seeking on-premise or edge AI solutions without compromising on advanced agentic and coding functionalities.

This release intensifies competition in the open-weight LLM space, pushing boundaries for parameter count, context window length, and on-device performance. Meta's focus on agentic capabilities and complex reasoning suggests a strategic pivot towards practical AI applications that can operate with significant autonomy. The limited commercial terms of its license will also shape how businesses and developers integrate Muse Spark Open, potentially fostering a vibrant ecosystem for customization and specialized deployments while Meta retains some control over widespread commercial exploitation.