深度拆解 Muse Glimmer，24GB 显存跑 30B Agent，Meta 到底做了什么？
Ranking
Overall
78
Content
90
Popularity
N/A
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Meta’s Muse Glimmer is an open 30B multimodal agent model designed for sustained local operation with 128K context on consumer GPUs. Its system-level combination of memory-efficient attention, quantization, agent-focused training, and speculative decoding aims to make private local agents more practical.
- GQA with two KV heads and alternating local/global attention sharply reduces long-context KV-cache requirements; 4-bit quantization enables a roughly 17GB variant targeting 24GB GPUs.
- A separate 1.8B vision encoder processes screenshots and documents, while runtime context management remains necessary to prevent visual-token growth and stale screenshots from overwhelming long tasks.
- On-policy distillation trains the model on states produced by its own rollouts, improving recovery from failed tool calls and deviations during multi-step workflows.
- The DFlash block-diffusion drafter accelerates speculative decoding; Meta reports RTX 5090 throughput rising from about 74.9 to 233.4 tokens/s for the 17GB quantized model, though results vary across agent benchmarks.
Sources (1)
深度拆解 Muse Glimmer，24GB 显存跑 30B Agent，Meta 到底做了什么？
Public signals
N/A
TL;DR - Meta’s Muse Glimmer is an open 30B multimodal agent model designed for sustained local operation with 128K context on consumer GPUs. Its system-level combination of memory-efficient attention, quantization, agent-focused training, and speculative decoding aims to make private local agents more practical.
- GQA with two KV heads and alternating local/global attention sharply reduces long-context KV-cache requirements; 4-bit quantization enables a roughly 17GB variant targeting 24GB GPUs.
- A separate 1.8B vision encoder processes screenshots and documents, while runtime context management remains necessary to prevent visual-token growth and stale screenshots from overwhelming long tasks.
- On-policy distillation trains the model on states produced by its own rollouts, improving recovery from failed tool calls and deviations during multi-step workflows.
- The DFlash block-diffusion drafter accelerates speculative decoding; Meta reports RTX 5090 throughput rising from about 74.9 to 233.4 tokens/s for the 17GB quantized model, though results vary across agent benchmarks.