vLLM Says It Reproduces DeepSeek Inference Performance with Gains in Both Throughput and Latency
On Oct 21, the vLLM project announced that, leveraging Wide-EP and EPLB, open-source vLLM can now reproduce DeepSeek's officially published DeepInfer inference performance — and even wins in low-latency scenarios. The result means DeepSeek-grade serving capability is no longer exclusive to the official stack, with direct reference value for the open-source inference ecosystem and inference-serving economics.
Confirmed
- Under DeepSeek's decode configuration (DeepSeek-V4.1-Flash, 131K context, nextn=3), vLLM reaches roughly 1M tokens/s/GPU throughput on GB300, within 6% of DeepSeek's official DeepInfer decoding performance.
- The corresponding revenue figure is $39 per GPU-hour.
- On latency: in real AgentX sessions, vLLM's dedicated decode path achieves single-user P90 TPS of 829 on 4 B300 GPUs, 26% faster than the GB300 standard decode path.
- The team thanked DeepSeek and related contributors in the post, and added data explaining the basis for the $39/GPU-hour calculation.
Why it matters
- DeepSeek's previously published high-throughput numbers were achieved mainly by its in-house stack, which third-party open-source stacks long struggled to fully replicate; vLLM's within-6% gap shows the open-source route is approaching official-level performance on top-tier hardware.
- It covers both high throughput (multi-user serving) and low latency (single-user experience), indicating a deployable serving capability rather than a single-metric optimization.
- The $39/GPU-hour revenue figure offers a concrete benchmark for the commercial viability of inference serving.
2026-10-21 ~ 2026-10-21 · 5 related posts
Primary sources
- vLLM says it can reproduce DeepSeek's inference performance, yielding $39/GPU/hr revenue — vllm_project ·
- vLLM matches DeepSeek's DeepInfer decode performance within 6%, generating $39/GPU/hr at off-peak pricing — vllm_project ·
- vLLM's specialized decode path hits 829 P90 TPS per user on four B300 GPUs, 26% faster — vllm_project ·
- [source] vLLM says it can reproduce DeepSeek's inference performance, yielding $39/GPU/hr revenue — vllm_project · 2026-10-21
- [source] vLLM matches DeepSeek's DeepInfer decode performance within 6%, generating $39/GPU/hr at off-peak pricing — vllm_project · 2026-10-21
- [source] vLLM's specialized decode path hits 829 P90 TPS per user on four B300 GPUs, 26% faster — vllm_project · 2026-10-21
- vLLM hits 829 P90 TPS single-user on four B300 GPUs, 26% faster than standard path on GB300 — vllm_project · 2026-10-21
1 near-duplicate retellings: zhuohan123