Unverified50% confidenceBenchmarkExact time
UP主实测64K上下文下生成速度每秒90多token，256K上下文降到每秒60多token
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/6/2026
First Seen
Valid until: 1/4/2027
Sources
Related Entities
Related Claims
UnverifiedUP主在4G显存电脑上实测MiniCPM-2B生成速度达每秒57个token，首字生成仅0.22秒78% similarUnverified56k拨号调制解调器的理论峰值约为每秒7KB74% similarUnverified云端API通常能达到每秒数十到上百个token的输出速度，而消费级GPU运行27B模型速度往往只有每秒10-30个token71% similarUnverifiedCollapse推理速度：CPU上单条10-token上下文约0.23毫秒，MPS在batch 1024下约230万token/秒，最近邻查询约0.48毫秒71% similarUnverified在4张B200上运行gpt-oss-120b纯预填充测试中，DWDP4在MNT 32K/ISL 32K配置下相比DEP4达到1.92倍加速，饱和状态下实现506K对329K tok/s（1.54倍）吞吐提升71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/979781API
curl https://kongchang.com/api/v1/knowledge/claims/979781MCP
get_claim(id=979781)