Unverified50% confidenceFactExact time
Mistral模型家族是SWA的典型代表，在每一层使用4096 token的窗口大小
1
Sources
50%
Confidence
Long-term
Relevance
8/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOllama的模型库包含数百个模型，包括Gemma 4、DeepSeek、Llama 4、Mistral等60% similarUnverifiedGQA将注意力头分组共享KV，已被LLaMA 3、Mistral等主流模型采用55% similarVerifiedMistral 7B的关键技术创新包括滑动窗口注意力机制（SWA）和分组查询注意力（GQA）54% similarUnverifiedMistral 推出了具身导航模型，参数量为 80 亿54% similarUnverifiedLangChain框架包含六大核心模块：Model I/O、Chain、Memory、Retrieval、Agent和Callbacks54% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/823645API
curl https://kongchang.com/api/v1/knowledge/claims/823645MCP
get_claim(id=823645)