Meta open-sources 816B multimodal Muse Spark Open on October 15, focused on agentic coding

Meta on October 15 open-sourced Muse Spark Open via Hugging Face, the full-size Muse series model that follows the smaller dense Muse Glimmer released in August and is currently the strongest US open-weights multimodal model. The model has roughly 816B total parameters, about 42B active per token, 62 layers, hidden dimension 8192, 256 experts per layer with 8 routed per token, and experts hold about 97% of total parameters. It supports text, image, audio and video inputs with unified text output and a maximum context of 1,000,000 tokens. The model is primarily trained for agentic coding scenarios, handling multi-file refactors, long debug sessions and tool calls, and supports the Harmony channel-style chat template with low, medium, high and xhigh reasoning effort tiers. The vision encoder Perception Encoder has 1.8B parameters; the audio encoder outputs one token every 80 ms, or 12.5 tokens per second. Checkpoints ship in BF16 (about 1.5 TB), NVFP4, GGUF and MLX, supporting Apple Silicon, 4-way GB300, 4-node 8-way H200 and single B200 deployments; the 2-bit MLX edition runs at around 18 tokens/s on a single Mac M3 Ultra with 512 GB of unified memory. Weights are published across the meta-models, mlx-community and ggml-org Hugging Face organisations.