GQA

LLMs Are Abandoning Full Attention—But Retrieval Reliability Remains Unproven

Open-weight models are shifting from full attention to hybrid and local designs, lowering long-context costs while raising retrieval risks.