Build with Muse Spark Open (dev.meta.ai)

Meta has announced the release of Muse Spark Open, an advanced open-weights model boasting 816 billion parameters and an innovative sparse mixture-of-experts architecture that activates around 42 billion parameters per token. This model is designed to handle diverse inputs—text, image, video, and audio—while supporting up to one million tokens of context. Its primary application is in creating autonomous agents capable of executing complex tasks such as reading extensive codebases, planning actions, and adapting based on feedback, making it a significant tool for AI/ML developers focusing on agentic workflows. The launch is impactful for the AI community as it ensures compatibility with numerous inference and training frameworks like SGLang, vLLM, and TensorRT-LLM right from day one, enabling wider accessibility and utility across various hardware platforms including NVIDIA, AMD, and Apple. Muse Spark Open supports advanced techniques like reinforcement learning and fine-tuning, offering customizable options through multiple quantization checkpoints. This model's launch heralds a new era of multimodal capabilities in AI, with potential applications spurring innovation and efficiency in AI-driven projects across industries.