What happened
DeepSeek's deployment note, with its English update on October 20, 2026, says the company is open-sourcing DeepInfer, the inference framework behind DeepSeek V4.1 Flash. The repository license is Apache 2.0. DeepSeek says one cluster currently serves 155 trillion tokens a day, with overall revenue of $38 per GPU-hour at valley pricing. The same note says KV cache management and load balancing are tightly coupled to DeepSeek's infrastructure and are not yet fully open-sourced. The README says the open release supports offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. It also says a complete service still needs a KV cache system, prefill and decode disaggregation, and load balancing, and that those pieces are not in this release. The note says the smallest efficient deployment it recommends is 64 GPUs of prefill plus 64 GPUs of decode.
Why it matters
Anyone who downloaded the October 17 Flash weights still needed a serving stack. DeepInfer is the stack DeepSeek says it runs, and the license on the repository is Apache 2.0. The number that makes the post look like a cost story, 155 trillion tokens a day and $38 per GPU-hour, describes DeepSeek's cluster. The note is explicit that the cache and the balancer behind that cluster are not in the open code. A reader can start from the offline example. A reader cannot reconstruct the quoted cluster from the repository alone.
The note's reference shape is a disaggregated vision-encoder, prefill, and decode system, plus storage nodes. It says prefill uses expert and data parallelism of 2 to 8, and decode uses 64 to 128. Tensor parallelism and context parallelism are not used for now, on that note.
The README says the default checkpoint is deepseek-ai/DeepSeek-V4.1-Flash-1017. It also says a KV cache kept for more than 72 hours was a design goal of the production system. That retention system is one of the parts the note says is not fully open.