What happens when 100 people call at once

Ten models released in 2026 — Qwen3.5 / 3.6 / 3.8, GLM-5.3, MiniMax-M3, DeepSeek-V4.1-Flash, gemma-4, granite-4.2 — measured against thirteen from 2024–25 on Hugging Face Inference Providers, at one concurrent user, then twenty-five, then a hundred.

Get InferGauge  → pip install infergauge Measured with infergauge.com

Read the full write-up  → Method, results, and how much the numbers move between runs

The short version

This section is written from the latest run every time the page loads, so it never disagrees with the table below it. If you are reading this sentence, scripting is off and the figures have not been filled in — the published CSVs linked under the table carry the same numbers.

How to read these numbers. Everything here is measured provider-side, from a datacenter, with one provider pinned per model. It describes how the provider behaves — not how your application will. Your network path, your prompts and your real traffic are not in this picture. For numbers that describe your own system, run InferGauge against your own endpoint.