➡️
继续阅读
-
Kimi K3 模型结构(5):宽度维度,Stable LatentMoE
Kimi K3模型采用Stable LatentMoE作为FFN,基于NVIDIA LatentMoE改进,将输入压缩至半宽latent空间,增加RMSN...
-
Claude Fable 5.1 vs. Fable 5: On real work, I couldn’t tell them apart.
Anthropic launched Claude Fable 5.1 this week, calling it “our most advanced ...
-
Building trust in agentic RAG starts with evidence
Basic retrieval-augmented generation (RAG) follows a straightforward pattern....
-
Kimi K3 模型结构(3):序列维度,Gated MLA 与 3:1 混合
本文介绍Kimi K3模型结构:采用3:1混合,每四层含一层Gated MLA(无位置编码、满秩sigmoid门),其余为KDA层。MLA低秩压缩KV至5...
-
内容创作者失误
During Naomi Osaka's match against Anastasia Zakharova at this year's...
-
Kimi K3 模型结构(2):序列维度(下),chunkwise 并行与下界衰减
本文介绍Kimi K3模型序列维度的chunkwise并行算法:将序列分块,块间递推状态、块内用矩阵乘并行计算。K3将每步log-decay加−5下界,使...