deepseek-v4.1IMAGE
DeepSeek DeepInfer report: 155T tokens/day per cluster, $38 per GPU-hour
A Zhihu technical report on DeepInfer, the inference stack behind V4.1 Flash, covers prefill/decode hardware choices and KV-cache targets plus real cluster data: 155 trillion tokens per day and $38 revenue per GPU-hour at off-peak pricing.
meng shao
@shao__mengArchived 11 days agoOriginally posted 2026.10.21
meng shao @shao__mengOriginally posted 2026.10.21
DeepSeek 开源 DeepInfer 对应的技术报告，发布在知乎了，还是一贯的技术含量极高，非常多的内部细节都公开了，学不完，根本学不完！ 比如 DeepSeek V4.1 Flash 设计之初的目标中，对 Prefill 和 Decode 的硬件选型和技术判断、KVCache 目标等，甚至还有单一集群的实际数据： 目前单一集群的服务吞吐是 155 万亿 Tokens/天，整体收入是每卡每小时 38 美金（按谷价）。 https://zhuanlan.zhihu.com/p/2095560295930508953
This post quotes another post: x.com/shao__meng/status/2112478119814947218
x.com/shao__meng/status/2112740048299082127Archived 2026.10.21 12:10