Unverified50% confidencePredictionExact time
UP主估计使用50GB显存显卡后推理速度可能提升到30~40 tokens/s（未经实测确认）
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/11/2026
First Seen
Valid until: 12/10/2026
Sources
Related Entities
Related Claims
Unverified该配置下35B模型实测单人最高速度约80 Tokens/秒，并发可达500 Tokens/秒以上74% similarUnverified本地部署DS4模型在STM32嵌入式开发测试中平均输出速度约为23 tokens/秒，推理时内存占用飙升至110GB，峰值速度达83.8 tokens/秒70% similarUnverified在预填充速度测试中，Halo约650 tokens/s，Spark约2000 tokens/s，Spark快了将近三倍70% similarUnverified作者认为12-30 token/s是舒适区，超过30 token/s算好；48-64GB的M5 Max约能跑到20 token/s68% similarUnverified200K token的上下文请求相比8K token请求，需要约25倍的KV-Cache显存68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/907339API
curl https://kongchang.com/api/v1/knowledge/claims/907339MCP
get_claim(id=907339)