Meta ships Muse Glimmer-30B as local agent model on consumer GPUs

A quiet upload from Meta's Superintelligence Lab is more interesting than the filename suggests. According to the model card on Hugging Face, Muse Glimmer-30B is a roughly 29.6B parameter dense causal transformer paired with a ~1.8B ViT-G/14 perception encoder, released under Apache 2.0 and shipped in Meta's ExecuTorch PTE format for on-device use.

The pitch is not raw capability against frontier hosted models, but autonomous agent work that runs on a single consumer GPU. Meta lists two 4-bit quantised variants: a K-Quant-Dynamic build that targets 32GB of VRAM with a claimed 0.2% degradation, and a 17GB K-Quant that fits a 24GB card at a claimed 1.0% degradation. A speculative-decoding drafter called DFlash reportedly takes an Nvidia RTX 5090 from 74.9 to 233.4 tokens per second, a 3.1x speedup, with smaller lifts on Apple's M4 Max (23.7 to 37.8 tok/s) and M5 Max (26.6 to 50.2 tok/s).

The self-reported benchmarks lean into the agentic angle: 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 51.2 on SWE-Bench Pro, plus 94.7 on AIME 2026 and 83.5 on GPQA Diamond. Meta positions the model against ones it calls 'Gemma4' and 'Qwen3.6' rather than any Llama entry of its own, which is a conspicuous omission on a card published by Meta.

Every number here is vendor-published on the Hugging Face card. There is no independent evaluation attached, the training data is not described, and the preparedness section resolves to 'Moderate or lower risk' across Chem/Bio, Cyber, and Loss of Control without publishing red-team detail. Anyone about to hand a local agent tool-use and file-system access should treat those labels as a starting point, not a finding.

Still, this class of release shifts the local-agent calculation. A 30B multimodal model that fits a single 24GB card under Apache 2.0 puts serious tool-using workflows within reach of startups, students, and any team whose data cannot leave the box.