DeepSeek publishes V4.1 Flash inference serving design with throughput and cost figures

DeepSeek has published the design of its V4.1 Flash inference service, according to a Zhihu post that Sophia Yang reposted on X. The post says one cluster handles 155 trillion tokens per day and earns $38 in revenue per card per hour at off-peak prices.

The design covers multi-protocol API adaptation, special-token handling, load balancing, and a KVCache backed by 3FS. It also includes speculative decoding and asynchronous scheduling. The figures come from DeepSeek's own post, and no independent verification is reported.