Muse Glimmer 30B: An On-Device Agentic Model
The Muse Glimmer team at Meta Superintelligence Labs releases Muse Glimmer, an Apache-2.0 30B multimodal model distilled from Meta's frontier model Muse Spark and built to run agentic workloads inside a 32 GB VRAM budget on a single consumer GPU.
Ask this paper
Design constraint. The architecture was chosen by training scaled ladders of candidate configurations under a fixed 32 GB VRAM ceiling. The released model is quantized to about 4 bits and uses a Dflash speculative-decoding head for single RTX 5090 inference.
Agentic results. Against Gemma4-31B and Qwen3.6-27B in thinking mode, it scores 75.5 vs 54.2 and 62.5 on MCPAtlas, 51.2 vs 36.9 and 50.2 on SWE-Bench Pro, and 47.6 vs 37.6 and 43.2 on WildClawBench.
Where it trails. Qwen3.6-27B is ahead on OSWorld-Verified (75.6 vs 65.9), SWE-Bench Verified (77.2 vs 76.0) and TerminalBench 2.1 (60.7 vs 51.7), so the lead is concentrated in tool use and long-horizon task benchmarks.
Scope. Text and image input, more than a hundred language varieties, and memory and privacy evaluations are part of the scorecard, with contamination screening reported for all benchmarks.
Why it matters. It is the first open-weight agent model from Meta's superintelligence group and a direct entry in the local-agent category, where most capable agent stacks still depend on cloud models.
Abstract
We introduce Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Muse Glimmer integrates multi-step reasoning, reliable tool use and failure recovery into a single model that runs locally on consumer devices without requiring cloud infrastructure. The model accepts text and image inputs and produces text outputs, was trained on data from over a hundred languages, and is quantized to approximately 4-bit precision for efficient single-GPU inference. Muse Glimmer achieves state-of-the-art performance for its size class on agentic benchmarks, setting a new standard compared to the models of its size: Gemma4-31B and Qwen3.6-27B on tool use, multi-step planning, and agentic task completion. We release the model under the Apache 2.0 license to foster open-source development of capable, safe, local AI agents.