Daily Tech Now

Daily News on AI, Big Tech, Linux, Security & Gadgets

Meta Unveils Muse Spark Open: A Revolutionary Open-Weights AI Model

Meta Unveils Muse Spark Open: A Revolutionary Open-Weights AI Model On October 15, Meta officially unveiled Muse Spark Open, a groundbreaking open-weights artificial intelligence model. According to the company, this release represents the most capable open-weights AI model developed in the United States to date. Advanced Multimodal Architecture Boasting a native multimodal architecture, the model…

Meta Muse Spark Open AI model multimodal architecture 816B parameters

Meta Unveils Muse Spark Open: A Revolutionary Open-Weights AI Model

On October 15, Meta officially unveiled Muse Spark Open, a groundbreaking open-weights artificial intelligence model. According to the company, this release represents the most capable open-weights AI model developed in the United States to date.

Advanced Multimodal Architecture

Boasting a native multimodal architecture, the model features a staggering 816 billion total parameters. Out of these, 42 billion act as active parameters. Furthermore, it supports an expansive one-million-token context window.

Engineers designed this system specifically for code generation and agent collaboration. Consequently, it demonstrates exceptional performance in handling long-horizon tasks, executing complex software engineering, and performing multi-step reasoning.

Desktop Optimization and Local Execution

Thanks to the latest MLX framework, this innovation becomes the first model in its class to run locally on a desktop. Amazingly, it requires only two Apple Mac Studio workstations equipped with M5 Ultra chips and 512 gigabytes of unified memory.

Starting today, developers can access the model via various API platforms. These include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. In addition, the model weights are now publicly available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

Seamless Integration and Rigorous Security

The architecture already supports major inference engines like SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Moreover, it features deep optimizations for silicon from AMD, Apple, Intel, and Nvidia. Developers can also easily customize and build upon the foundation using the PyTorch framework.

During the post-training phase, researchers specifically reinforced the safety alignment of the model. As a result, benchmark tests indicate that the system maintains strict safety guardrails across all tested risk dimensions.

Finally, Meta released the model weights under the Muse Model License. Although this license includes restrictive commercial terms, the vast majority of developers can still acquire and utilize this powerful tool completely free of charge.

About the Author

Trang Nguyen Avatar

Leave a Reply

Your email address will not be published. Required fields are marked *