What happened?
Meta Muse Spark Open arrives as Meta's new open-weight model. The company launched Muse Spark Open, its most capable open-weight model to date. With 816 billion total parameters (42 billion active per token) and a 1 million token context window, the model is natively multimodal and optimized for agentic coding tasks, complex software engineering, and multi-step reasoning. It can run on two Mac Studios with the M5 Ultra chip using the MLX framework, creating a desktop supercomputer.
Weights are available on Hugging Face, Amazon SageMaker, Google Cloud, and Microsoft Foundry. Support is also available via partner APIs such as Together AI, Fireworks, and Groq. Meta claims it is the most advanced US open-weight model, competing with Chinese leaders like Qwen and Kimi.
Why does it matter?
Open-weight models allow local customization and reduce dependence on closed APIs. Muse Spark Open brings long-horizon agent capabilities to the desktop, especially in the Apple ecosystem. This accelerates AI tool development for programmers and researchers who prefer full data control.
Optimization for Apple Silicon highlights an unusual partnership and makes massive model inference more accessible without high-cost NVIDIA hardware.
What changes in practice?
Developers can download the weights, fine-tune with PyTorch/TorchTitan, and run coding agents locally on high-spec Macs. The model includes extra safeguards for open use and supports frameworks like SGLang and vLLM. License is the Muse Model License with limited commercial terms.
Source: official Meta AI Research blog. Official announcement.
Image credit: Meta — Source: X/@AIatMeta
By GeekikiBot
