Unverified50% confidenceBenchmarkExact time

同一模型在DeepCWE上因智能体框架不同得分从65.5到74.2不等，官方harness用的是区间上限，Claude Code的harness下只有69.8

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

9/27/2026

First Seen

Valid until: 12/26/2026

Sources

DeepSeek V4.1 Flash实测：效率革命还是精心挑选的图表？

Related Entities

DeepSeek V4.1 Flashmodel DeepCWEconcept

Related Claims

Unverified同一个DeepSeek v1.1模型在Terminal Bench 2.0上不同Harness下的得分从69.8%跨到74.2%69% similar UnverifiedMoE架构中Mixtral 8x7B模型总参数46.7B,但每个token仅激活约12.9B参数(2个专家),推理成本接近13B参数的稠密模型,性能可媲美70B级别62% similar Verified在 DeepSWE 基准上，GPT-5.6 各推理档位得分分别为 Low 45 分、Medium 61 分、High 69 分、X-High 71 分、Max 73 分62% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/954587

API

curl https://kongchang.com/api/v1/knowledge/claims/954587

MCP

get_claim(id=954587)