Benchmarking DeepSeek-R1 and Qwen3 on EU Infrastructure

We ran a series of standardized benchmarks on our Amsterdam cluster to validate latency and throughput for two of the most requested models: DeepSeek-R1 and Qwen3-235B.

Setup

Tests were run from Frankfurt with vLLM on NVIDIA H200 GPUs. We measured time-to-first-token, tokens per second, and end-to-end latency for prompts ranging from 1K to 128K tokens.

Results

DeepSeek-R1 averaged 42ms TTFB for short prompts and scaled linearly to 180ms at 128K context. Qwen3-235B showed slightly lower TTFB but higher memory usage at long context.

Both models comfortably exceeded 100 tokens per second for generation, with batching improving throughput by up to 3x.