Meta's Muse Spark Open release

Meta released Muse Spark Open, an 816B-parameter open-weight model on Hugging Face as BF16 and NVFP4 checkpoints under the Muse Model License. The model uses a Mixture-of-Experts Transformer with a Perception Encoder, 42B active parameters at inference, a 1 million token context window, and accepts text, image, video, and audio inputs. It targets coding and agentic workloads and is available through Azure AI Foundry and hosted inference providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

The open release shares the same base model as the proprietary Muse Spark 1.3, with additional safety post-training added before open release, reversing Meta's closed-weight strategy from April. Apple optimized the model for its MLX framework; running the full model requires two linked M5 Ultra Mac Studios. The 4-bit NVFP4 quantization needs at least 700GB of GPU memory, while a 3-bit version tested on a 512GB M3 Ultra consumed roughly 344GB and ran at about 21 tokens per second on short prompts. Meta's Helion GPU kernels report kernel-level speedups of 1.03 to 3.95 times, but the end-to-end improvement on AMD MI350 hardware is 3%.

The Muse Model License is custom and not OSI-approved. All licensees must display a visible 'Built with Muse Spark Open' notice. Only model-as-a-service providers whose revenue exceeds $20 million in any twelve-month period must obtain a separate agreement with Meta, with that threshold covering the licensee and affiliates. Internal use, academic research, and products that embed the model as a feature are exempt. Hugging Face access requires submitting full legal name, date of birth, and full organization name with corporate identifiers.