Unverified50% confidenceBenchmarkExact time

在实验中，由4B模型编写技能时27B模型在LiveMath上得分61.0%，而27B自己编写技能时得分56.3%

1

Sources

50%

Confidence

Long-term

Relevance

8/31/2026

First Seen

Sources

WikiSkill论文解读：4B小模型为何能当27B大模型的好老师

Related Entities

Qwen-3.5-4Bmodel Qwen-3.6-27Bmodel LiveMathconcept

Related Claims

Unverified混合精度训练中，FP16 每个参数占 2 字节显存，FP32 每个参数占 4 字节显存63% similar Unverified2.4万亿参数模型在BF16精度下考虑KV Cache和激活值后实际约需1.4-1.5TB显存62% similar UnverifiedMoE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62% similar Unverified早期Codex在HumanEval基准测试中以37%的准确率独立解决编程题，达到当时最高水平62% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/828181

API

curl https://kongchang.com/api/v1/knowledge/claims/828181

MCP

get_claim(id=828181)