Meta releases 30B Muse Glimmer for local AI agents
Meta has released Muse Glimmer, a 30-billion-parameter multimodal model distilled from Muse Spark. The model is designed for continuously running local agent workflows and is distributed under the Apache 2.0 license. Its inputs are text and images, while audio is unsupported and video is handled as separate frames. The context length exceeds 131,072 tokens, and the stated knowledge cutoff is January 4, 2026.
Meta reduces the model to approximately 4-bit weights, bringing the language model below 20 GB. The K-Quant-17GB build targets 24 GB of VRAM with a reported 1.0% average degradation across 15 benchmarks; another build targets 32 GB. The model can run on a single consumer GPU or an M4/M5 Max Mac without a network call. Hugging Face provides BF16, GGUF, ExecuTorch and DFlash components.
DFlash predicts 16-token blocks for parallel verification. Meta measured throughput rising from 74.9 to 233.4 tokens per second on an RTX 5090, a 3.1-fold increase. Muse Glimmer led listed comparisons on MCP Atlas, DeepSearch QA and SWE-Bench Pro, but trailed Qwen3.6-27B on OSWorld-Verified, TerminalBench 2.1 and SWE-Bench Verified. Meta rated its chem/bio, cyber and loss-of-control risks as moderate or lower.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.