Muse Glimmer

Muse Glimmer, released on August 10, 2026 under the Apache 2.0 license, is Meta's first fully open-weight model since it moved on from the Llama family. It is a dense transformer with roughly 29.6 billion parameters in total, 52 language-model layers, and a 1.8 billion parameter ViT-G/14 perception encoder, trained on Muse Spark's outputs through logit distillation during pretraining. Input is interleaved text and images, output is text, the training data spans more than 100 languages, and the model card lists a context length of 131,072 tokens or more and a January 4, 2026 knowledge cutoff. Meta trained it around the loop an autonomous agent actually runs: form a plan, call tools against precise schemas, interpret the results, keep going, and diagnose and retry when a tool fails. Reasoning effort is controllable to trade quality against speed. Meta reports 51.2 on SWE-Bench Pro, 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 65.9 on OSWorld-Verified, and 94.7 on AIME 2026, and positions it against Gemma 4 31B and Qwen3.6 27B. Sized for 24 GB and 32 GB consumer GPUs and validated on M4 Max and M5 Max MacBooks and an RTX 5090, Muse Glimmer suits local coding agents, computer-use experiments, and automation where the whole agent runs on one machine.