DeepSeek open-sources DeepInfer, its internal inference stack

DeepInfer exposes the serving techniques behind DeepSeek's very low prices. Analysts estimate that the stack is a key reason DeepSeek can sustain high margins despite aggressive per-token pricing.

Developers are benchmarking it against vLLM and SGLang, and many are excited about self-hosting DeepSeek models more cheaply.

If the claims hold, the release pressures inference-hosting startups and could compress per-token prices across the open-model market.

The caveat is that the stack is presumably tuned to DeepSeek's own architectures and hardware, so gains may not carry over to other models.