Muse Spark Open 816B A42B released

This repository holds the full-precision weights of Muse Spark Open, the open-weight model Meta announced as the most capable US open-weight model to date (Meta's release post). The model card describes a natively multimodal model built for coding and agent work: it reads text, images, video and audio, writes text, and accepts up to 1M tokens of context (model card).

The two numbers in the name explain the hardware picture. Muse Spark Open is a mixture-of-experts model: each layer has 256 small sub-networks, called experts, and a router sends each token through only 8 of them, so roughly 42B of the 816B parameters do the work for any given token. Hugging Face's launch post spells out the catch: all the experts still have to sit in memory, so the model needs the memory of an 816B model while running at the speed of a 42B one (Hugging Face). In this BF16 form the checkpoint is about 1.5 TB, the card asks for at least 2TB of GPU memory, and Hugging Face's transformers example spreads it across four nodes of eight H200 GPUs. A separate NVFP4 checkpoint exists for smaller deployments.

Open weights here does not mean an unrestricted download. The repository is gated: you request access and accept the Muse Model License before getting the files. That license allows use, modification, fine-tuning and redistribution, but requires each product or service built on the model or its derivatives to display "Built with Muse Spark Open" in its interface, documentation, website or settings. It also covers companies that offer AI models to others through an API: if the company's gross revenue, together with its affiliates, exceeds $20 million in any 12-month period, it must notify Meta and sign a separate agreement before offering this model that way. Internal use, research, and using the model to power features of a company's own non-API product are excluded from that clause. Meta's own post describes these as limited commercial terms (Meta).