816B total / 42B active MoE (approximately 822.0B serialized BF16 parameters)
1M tokens
Muse Model License
Meta's unnumbered open-weight Muse Spark release for long-horizon agentic work, coding, tool use, multimodal reasoning, synthetic-data generation, and LLM-as-a-judge evaluation. Muse Spark Open is a 62-layer sparse Mixture-of-Experts Transformer with an 8,192-dimensional hidden state, gated attention with QK normalization, 128 query heads, 16 key-value heads, a 202,048-token vocabulary, and 256 routed experts selecting eight per token. Meta advertises 816B total and 42B active parameters, while Hugging Face inventories approximately 822.0B serialized parameters in the BF16 checkpoint. A separate approximately 1.8B-parameter, 50-layer perception encoder with a 1,536-dimensional hidden state, 16 attention heads, 14-by-14 patches, sparse attention, learned absolute positions, and 2D RoPE supports text, images, documents, audio, and video as input with text output. The model supports a 1M-token context window, more than 100 training languages, multi-step planning, sequential tool invocation, failure recovery, coding across large repositories, web and computer use, subagent coordination, and low, medium, high, and xhigh reasoning effort. Meta describes this as a distinct open release of Muse Spark with additional safety post-training intended to improve robustness when deployed without the system-level guardrails surrounding hosted Muse checkpoints. Official BF16 and NVFP4 checkpoints are available through manually gated Hugging Face repositories; Meta states that BF16 deployment requires approximately 2TB of GPU memory and NVFP4 approximately 700GB. Supported launch runtimes include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth, with verified NVIDIA, AMD, Intel, and Apple deployments; Meta and Apple demonstrated local operation across connected M5 Ultra Mac Studios. Meta reports 61.8 on JobBench, 89.2 on MCP-Atlas, 80.4 on OSWorld Verified, 26.1 on HLE-Diamond without tools, 58.5 on SciCode, 89.0 on GPQA Diamond, 60 on DeepSWE 1.1, 76.4 on Terminal-Bench 2.1, 50.0 on ZeroBench with tools, 73.6 on SimpleVQA with tools, and 61.5 on VideoMMEv2 at xhigh reasoning. The custom Muse Model License allows use, modification, redistribution, derivatives, and commercial use subject to its usage policy and attribution requirements, while qualifying Model-as-a-Service businesses whose affiliated revenue exceeds $20 million must notify Meta and enter a separate written agreement.