Unverified50% confidenceFactExact time
GQA compresses the KV Cache by about 8x relative to MHA and has become the standard choice for mainstream open-source models such as Llama 2/3, Qwen, and Mistral.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCompared to MHA, GQA can compress the KV cache size to one-quarter without noticeably degrading model quality.73% similarUnverifiedThe 70B version of LLaMA 2 uses 8-group GQA, reducing the KV cache to 1/4 of MHA while barely affecting model quality.70% similarUnverifiedMLA theoretically reduces the GPU memory footprint of KV Cache to 5% to 13% of standard MHA by compressing KV matrices into low-dimensional latent vector caches.69% similarUnverifiedDeepSeek-V2 论文显示 MLA 可将 KV Cache 显存降低至标准 MHA 的约5%至13%，同时维持相近性能65% similarUnverifiedQuantization formats such as GGUF and AWQ can compress models with tens of billions of parameters to a size runnable on consumer-grade GPUs while maintaining near-original performance.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1025447API
curl https://kongchang.com/api/v1/knowledge/claims/1025447MCP
get_claim(id=1025447)