vLLM reports DeepSeek-V4.1-Flash performance on AgentX with real-world agentic workloads

What happened

vLLM reports 954K tokens/s/GPU total throughput (cached+input+output) at 196 tokens/s P90 per user on SemiAnalysis AgentX with real-world agentic workloads. At DeepSeek's off-peak pricing, vLLM says this delivers $38.96/GPU-hour revenue and calls it the best DeepSeek-V4.1-Flash result on AgentX ever. Its engineering blog explains the optimizations and how to reproduce the results: https://vllm.ai/blog/2026-10-23-deepseek-v41-flash-pd-wide-ep

Read next

• More changes in today’s briefingMoonshot launches Kimi K3.1 Pro with weights promised November 4 and faster Highspeed API

The story so far

vLLM reported DeepSeek-V4.1-Flash throughput of 954K tokens/s/GPU, counting cached, input and output tokens, with P90 per-user throughput of 196 tokens/s on SemiAnalysis AgentX using real-world agentic workloads. vLLM called it the best DeepSeek-V4.1-Flash result ever on AgentX. Applying DeepSeek’s off-peak pricing, vLLM said the result translates to $38.96/GPU-hour in revenue. Its engineering blog describes the optimizations and provides instructions for reproducing the benchmark results.