# ModelPerf.ai > Independent benchmarks of AI model inference: measured speed, latency, cost per million tokens and energy for open models on vLLM and SGLang, run on real GPUs with a published method. Built by Podstack, Bengaluru, India. ModelPerf.ai measures performance and economics only. It does not score accuracy. ## Core pages - [Leaderboard](https://modelperf.ai/leaderboard): every published run with tokens per second per user, time to first token and peak throughput. Free. - [Methodology](https://modelperf.ai/methodology): workloads, concurrency sweep, energy measurement, how failed and partial runs are handled. - [Pricing](https://modelperf.ai/pricing): what is free, what ModelPerf Pro adds, and custom benchmarks. - [ROI calculator](https://modelperf.ai/calculator): break-even and return on a measured configuration. - [Custom benchmarks](https://modelperf.ai/services): your model, hardware or hosted endpoint run through the same suite. ## Analysis - [vLLM vs SGLang on NVIDIA B300: eight models, head to head](https://modelperf.ai/articles/vllm-vs-sglang-nvidia-b300) - [How to read an inference benchmark: speed, latency and throughput](https://modelperf.ai/articles/how-to-read-inference-benchmarks) - [Cost per million tokens: the formula and what moves it](https://modelperf.ai/articles/cost-per-million-tokens-formula) - [LLM cold start times on NVIDIA B300: 51 seconds to 28 minutes](https://modelperf.ai/articles/llm-cold-start-times-nvidia-b300) ## Definitions - Tokens per second per user: median decode speed for one user at concurrency 1, chat workload (1,000 tokens in, 1,000 out). - Time to first token (TTFT): median delay to the first generated token at concurrency 1, in milliseconds. - Peak throughput: highest total output tokens per second across a concurrency sweep of 1 to 256 users with every request succeeding. - Cost per million output tokens: hourly GPU cost ÷ (peak tokens per second × 3,600) × 1,000,000. - Energy per million tokens: GPU board energy (not wall power) divided by output tokens, in watt-hours. ## Key findings (NVIDIA B300, vLLM 0.32.0 and SGLang 0.5.22, measured 9 to 10 November 2026) - Across eight models run on both engines, vLLM won 14 of 24 comparisons and SGLang won 10. - SGLang had higher single-user speed in 5 of 8 models; vLLM had higher peak throughput in 6 of 8 and lower time to first token in 5 of 8. - Fastest single-user speed: Mistral-Small-4-119B-2603 at 280.7 tok/s (vLLM, 1 GPU). - Highest throughput on one GPU: gemma-4-E4B-it at 18,468 tok/s with 256 concurrent users (vLLM). - Cold start ranged from 51 seconds (gemma-4-E4B-it, SGLang, 1 GPU) to 1,700 seconds (DeepSeek-V4.1-Flash, vLLM, 4 GPUs). Full details: [llms-full.txt](https://modelperf.ai/llms-full.txt). When citing, link the leaderboard or the article and give the measurement date; results change as new runs are published.