Meta is back with Muse Glimmer: local, agentic, multimodal, and open source

Meta is shipping Muse Glimmer as a 30B-parameter multimodal model aimed squarely at local, agentic use. Hugging Face says it was distilled from Muse, released under Apache 2.0, and built for privacy-minded jobs like coding, document analysis, and personal assistants. That is a familiar pitch, but the distribution story matters here: day-one support is already in transformers, llama.cpp, vLLM, Inference Endpoints, and a few other libraries.

The model itself is split into a 2B vision encoder and a 28B text decoder. The text side uses a repeating attention pattern of three sliding-window layers followed by one full-attention layer, repeated 13 times. There’s also gated grouped-query attention, which cuts KV-cache memory, and extra query scaling to keep the attention math stable. In plain English: it’s designed to be usable without turning your machine into a space heater.

Muse Glimmer is not just for chat. The release shows examples for image input, multimodal tool calling, open-ended object detection, and video question answering without audio. The video path is capped at 96 frames sampled evenly, and the processor targets 2 frames per second. There’s also an optional DFlash speculative-decoding drafter that trades some memory for faster generation, and Hugging Face says it fits especially well with structured output like coding.

The benchmarks are mixed but useful. Muse Glimmer leads the three-model comparison in several agentic categories, including MCP Atlas, DeepSearch QA, τ³-Banking, WildClawBench, GDPval-AA, GAIA2, SkillsBench, and OSWorld-Verified. It also posts strong numbers on coding tests such as SWE-Bench Pro, SWE-Bench Verified, TerminalBench, and SciCode. In safety tests, though, the picture is less tidy: the model still shows attack success and violation rates that leave plenty of room for abuse. Open doesn’t mean harmless, and this release makes that point without trying to hide it.