Unverified60% confidenceFactExact time
盘古2.0 Flash采用MoE（混合专家）架构，参数规模为92B A6B（总参数92B，激活参数6B）
2
Sources
60%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
华为盘古2.0 Flash高考实测：613分成绩与稳定性深度解析
bilibiliwycode_cn7/7/2026
Related Claims
Unverified盘古2.0 Flash在GPQA Diamond科学推理测试中（Thinking模式）得分83.7，低于千问3.6 35B的86、27B稠密模型的87.8和DeepSeek V4 Flash的88.175% similarVerifiedDeepSeek V4 Flash uses a Mixture of Experts (MOE) architecture with a total parameter count exceeding 600 billion (600B+).71% similarUnverified阶跃星辰（StepFun）发布了开源模型Step 3.5 Flash，拥有1960亿总参数，运行时只激活110亿参数67% similarUnverified在FPS游戏生成测试中，V4 Flash生成988行代码，V4 Pro生成1456行代码66% similarUnverifiedFlash Attention 2 在 A100 上实现约 2 倍加速，Flash Attention 3 针对 Hopper 架构（H100）引入异步执行与低精度 FP8 支持65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/330493API
curl https://kongchang.com/api/v1/knowledge/claims/330493MCP
get_claim(id=330493)