Meta releases Muse Glimmer, a 30B-parameter open-weight agent model engineered to run on a single 24GB GPU.
The AMW Read
Meta ships a benchmarked, Apache-2.0 30B agent model with explicit KV-cache, quantization, and speculative-decoding tradeoffs for 24GB local deployment, meaningfully advancing open-weight on-device agent architecture without introducing a new top-tier lab or resolving a named debate.
Meta releases Muse Glimmer, a 30B-parameter open-weight agent model engineered to run on a single 24GB GPU.
Meta released Muse Glimmer, a roughly 30B-parameter multimodal agent model with 128K-token context, under an Apache 2.0 license. It ships as two 4-bit quantized builds — about 16.8GB for 24GB-VRAM cards and about 20GB for 32GB cards — plus a 1.8B-parameter ViT-G/14 vision encoder and a DFlash inference-acceleration module, deployable via llama.cpp, MLX, and ExecuTorch. The architecture pairs grouped-query attention (32 query heads, 2 KV heads) with a hybrid local/global attention pattern — 39 of 52 layers use a 2048-token local window, only 13 layers carry full context — cutting estimated KV-cache size from roughly 20GB under standard multi-head attention to about 1.7GB. Meta reports 0.2%-1.0% average accuracy loss across 15 benchmarks and a 3.1x decode speedup from DFlash's parallel block-diffusion drafter, versus a standard one-token-at-a-time drafter.
The release is notable less for parameter count than for what it optimizes: an agent loop that keeps accumulating tool outputs, code logs, and screenshots across dozens of steps, not a single chat turn. By engineering KV-cache and decode cost to a specific 24GB hardware target and licensing it permissively, Meta pushes multi-step, tool-using, screen-reading automation toward hardware that doesn't need a hosted API call per step.
Builders get an open, benchmarked base tuned for screenshot-heavy, multi-step, tool-calling work rather than a repurposed chat model, with a low/medium/high/xhigh reasoning-budget dial for trading cost against task success. For on-device and edge-AI infrastructure investors, the differentiator is architecture — attention pattern, KV-cache footprint, speculative decoding — not raw parameter count.