Unverified50% confidenceFactExact time
Muse Glimmer是稠密模型，每次推理激活全部300亿参数，Flash Attention可将速度提升约1.5到3倍
1
Sources
50%
Confidence
Long-term
Relevance
8/19/2026
First Seen
Sources
Related Claims
UnverifiedMuse Glimmer采用混合注意力机制：大部分层只关注2048 token的局部窗口，每隔四层做一次全局全注意力77% similarUnverifiedIn real-world testing, Muse Glimmer's visual capabilities stood out, achieving perfect scores in both advanced image recognition and invoice recognition, with visual capabilities already on par with Qwen.74% similarPartially VerifiedMuse-Glimmer-30B is a multimodal model with image-text understanding capabilities.70% similarUnverifiedMeta发布的Llama 3系列采用分组查询注意力机制（GQA），在推理效率上较传统多头注意力有显著提升64% similarUnverifiedFlashAttention、滑动窗口注意力等工程优化技术使百万级上下文在可接受的计算资源范围内成为可能64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/772088API
curl https://kongchang.com/api/v1/knowledge/claims/772088MCP
get_claim(id=772088)