DownloadMuse Spark Open is Meta Superintelligence Labs' open-weight multimodal Mixture-of-Experts model designed for advanced coding, reasoning, and autonomous agent workflows. It contains 816B total parameters with 42B activated per token, using 256 experts and a dedicated 1.8B-parameter perception encoder. The model natively processes text, images, audio, and video and generates text, with a 1M-token context window for extended tasks. It supports more than 100 languages and configurable low, medium, high, and xhigh reasoning effort. Its capabilities include multi-step planning, tool calling, error recovery, software engineering, multimodal analysis, synthetic data generation, and model evaluation. Additional safety post-training improves resistance to misuse in standalone deployments. Released under Meta's Muse Model License, it supports BF16 and NVFP4 weights and deployment through vLLM, SGLang, TensorRT-LLM, and other inference frameworks.
Features
- 816B total parameters with 42B activated per token
- 256-expert Mixture-of-Experts Transformer architecture
- Dedicated 1.8B-parameter multimodal perception encoder
- Native text, image, audio, and video understanding
- 1M-token context window for extended agent workflows
- Low, medium, high, and xhigh configurable reasoning effort
- Advanced coding, tool use, planning, and failure recovery
- BF16 and NVFP4 weights with vLLM, SGLang, and TensorRT-LLM support
Project Samples
