Unverified50% confidenceBenchmarkExact time
该测试中预填充速度固定在约 500~600 token/s，首 token 耗时随输入长度线性增加
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/17/2026
First Seen
Valid until: 12/16/2026
Sources
Related Entities
Related Claims
Unverified在8GB显卡上，token生成速度约为4.2到5 token/秒，且随着上下文变长速度从约6 token/秒下降到4.2左右74% similarVerified在输出（Decode）速度测试中，2 万 token 以内可达约 45 tokens/秒，3 万 token 以上骤降至约 20 tokens/秒72% similarUnverified在案例测试中Gemma分析三页PDF财务报告耗时约13.78秒，生成1064个tokens，速度为每秒16 tokens72% similarUnverified该测试中模型短上下文场景输出速度约20 token/秒，200K+上下文降至约13 token/秒71% similarVerifiedHuman reading speed is roughly 5-8 tokens per second70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930285API
curl https://kongchang.com/api/v1/knowledge/claims/930285MCP
get_claim(id=930285)