Unverified50% confidenceSolutionExact time
The fix for Softmax Collapse is to use log-softmax and compute it stably via the log-sum-exp trick.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe bottleneck of on-device RAG is memory: model weights plus runtime overhead must be kept within a range acceptable to the application, so the 191MB text mode is easier to integrate than the 567MB full mode.52% similarUnverified稀疏激活技术通过条件计算机制仅激活与当前输入最相关的参数子集，可减少从HBM加载的权重数据量，缓解内存墙带宽瓶颈51% similarUnverifiedThe model catalog refresh added two providers, Boundless and Standard Compute; SambaNova is now fetched from the catalog, with MiniMax M3 as the default model.51% similarUnverifiedall-MiniLM-L6-v2 将文本压缩成 384 维稠密向量，模型体积约 80MB，底层基于 BERT 精简变体 MiniLM 通过知识蒸馏训练51% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088265API
curl https://kongchang.com/api/v1/knowledge/claims/1088265MCP
get_claim(id=1088265)