Unverified50% confidenceBenchmarkExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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69% similarUnverified2.4万亿参数模型在BF16精度下考虑KV Cache和激活值后实际约需1.4-1.5TB显存66% similarUnverified在严格格式评分测试中，TwIL-LM 的 1.7B 模型据称击败了体量是其 5 到 15 倍的模型65% similarUnverified在覆盖九大维度共403道题的测试中，Muse-Glimmer-30B综合均分为76.04，比Corpus 3.6 27B的75.57高出0.47分64% similarUnverifiedMoE模型中'35B'指总参数量，每个token只激活少量Expert（通常2-4个），实际计算量可能仅相当于6-7B的Dense模型64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942874API
curl https://kongchang.com/api/v1/knowledge/claims/942874MCP
get_claim(id=942874)