Infrastructure & deployment
DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Since Day 0
vLLM reports a 1.9× low-latency improvement and 5.3× throughput under a 150 TPS constraint on DeepSeek-V4.1-Flash's AgentX serving workload.
Why it matters: For deployment, replay and CUDA graphs need to be considered together: graphs help avoid launch overhead on short prompts. The serving configuration also matters—vLLM used TP4 with FlashInfer for low latency and DEP2 with DP attention for high throughput.