Unverified50% confidenceBenchmarkExact time
在同一套简洁提示词和参考重叠评分器下，已发布的406M专用模型在ROUGE-1上至少领先五个商用托管大模型6.2分
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/23/2026
First Seen
Valid until: 12/22/2026
Sources
Related Entities
Related Claims
Unverified微调后的406M模型在测试集上取得36.33 ROUGE-1，作者自建的1.2B系统为35.41 ROUGE-1，两者在统计意义上无法被区分69% similarUnverified在MoE模型中每个token只激活少量Expert（通常2-4个，Top-K路由），实际计算量可能仅相当于6-7B的Dense模型64% similarUnverifiedGLM4在SWE-Bench基准测试中得分46.2%，显著优于上一代模型64% similarUnverifiedGLM-5.2在DesignArena中以1360分夺得榜首64% similarUnverifiedGLM-5.3在六项评测中的成绩为：终端编程基准3.0得28.3、深度软件工程基准得66.9、智能体中场考试得28.5、自动化基准得48.2、带工具的人类终极考试得62.5、GDP价值评估AA2版得17.6963% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942880API
curl https://kongchang.com/api/v1/knowledge/claims/942880MCP
get_claim(id=942880)