Approximately 29.6B including approximately 1.8B perception encoder
131,072 tokens
Apache License 2.0
Open-weight multimodal agentic model distilled from Muse Spark and optimized by Meta for autonomous workflows on consumer hardware. Muse Glimmer uses a 52-layer dense causal Transformer with a 6,656-dimensional hidden state, SwiGLU feed-forward layers, 32 query heads and 2 key-value heads, gated attention, and a repeating three-local/one-global attention pattern; local layers use a 2,048-token sliding window, while the official configuration supports 131,072 tokens overall. A dedicated approximately 1.8B-parameter, 50-layer ViT-G/14 perception encoder supports interleaved text and image input with text output and up to 4,096 visual tokens per image. Meta trained the model for end-to-end agentic task completion, schema-based tool calling, multi-step reasoning, failure recovery, coding, and compatibility with agent scaffolds, while also supporting more than 100 languages and low, medium, high, and xhigh reasoning-strength controls. Meta reports scores including 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 51.2 on SWE-Bench Pro, 76.0 on SWE-Bench Verified, 77.0 on IFBench, and 94.7 on AIME 2026, positioning it strongly against Gemma4-31B and Qwen3.6-27B in its size class. The release includes BF16 weights, K-Quant-Dynamic and K-Quant-17GB GGUF variants, a frozen perception encoder, and a five-layer DFlash block-diffusion drafter that proposes 16-token blocks for speculative decoding. Meta says the approximately 4-bit quantizations reduce the language model below 20 GB with measured average benchmark degradation of 0.2% for K-Quant-Dynamic and 1.0% for K-Quant-17GB; with DFlash, it reports 233.4 tokens per second and a 3.1x speedup on an RTX 5090. Official artifacts target local and edge deployment through GGUF and ExecuTorch, with integrations announced for llama.cpp, MLX, vLLM, SGLang, Ollama, LM Studio, and Unsloth.