Meta Releases Muse Glimmer, a 30B Open Agent Model That Runs on One Consumer GPU
Meta Superintelligence Labs shipped an Apache 2.0 model built for always-on local agents, fitting on 24GB of VRAM and beating Gemma4-31B by 21 points on agentic orchestration benchmarks.
Meta Superintelligence Labs released Muse Glimmer on August 10 — a 30-billion-parameter open-weight model built specifically for always-on AI agents running on a single consumer GPU, no cloud connection required. It’s free under Apache 2.0, and the weights are live on Hugging Face in BF16, GGUF k-quant, and ExecuTorch builds.
The model is a dense causal transformer with a dedicated 1.8B-parameter perception encoder bolted on, distilled down from Meta’s closed frontier model, Muse Spark. At full precision it needs 55GB of memory, but Meta’s K-Quant-17GB configuration drops that to 24GB of VRAM — one RTX 5090, one Mac Studio, one workstation GPU — while losing just 1.0% accuracy. A more conservative K-Quant-Dynamic mode holds degradation to 0.2% at 32GB. The real trick is DFlash block-diffusion decoding, which predicts 16 tokens per forward pass instead of one, delivering a 3.1x speedup — 74.9 to 233.4 tokens per second on an RTX 5090.
Meta benchmarked it against the two closest open-weight competitors in its size class: Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B. On MCP Atlas, an agentic tool-orchestration benchmark, Muse Glimmer scored 75.5 against Gemma4’s 54.2 — a 21-point gap. It’s not a clean sweep, though. On OSWorld-Verified, a computer-use benchmark, Qwen3.6-27B leads at 75.6 versus Muse Glimmer’s 65.9, and Qwen also wins TerminalBench 2.1. The model is purpose-built for agentic orchestration and function-calling, not general computer control — and the benchmarks bear that specialization out.
This is Meta’s clearest bet yet that the next wave of AI agents runs locally rather than through an API call. A coding assistant, a browser agent, or a model-based evaluator that never leaves your machine sidesteps latency, per-token cost, and the privacy tradeoffs of routing every action through a hosted frontier model. Apache 2.0 licensing means startups can fine-tune and ship it commercially with no royalty string attached — the same playbook that made Llama the default base model for a huge share of the open-source AI ecosystem.
The open-weight tier at 27-31B parameters is now genuinely contested — Meta, Google, and Alibaba all fielding entries within weeks of each other — which is good news for anyone building on top of it and bad news for closed-API vendors selling inference at that capability level.