# IQ-Server

> IQ-Server is a streaming server built for one job: carrying the tokens of an LLM answer from your application backend to your users’ browsers, in real time, over Server-Sent Events.

- Your backend publishes each chunk of model output over a plain HTTP request, one JSON line per event, from any language: PHP, Python, Node.js, Go, Ruby, Java and .NET, or anything that speaks HTTP.
- Browsers read the stream with the built-in EventSource, with no SDK. IQ relays events exactly as sent, so it works with AG-UI, the Vercel AI SDK, plain text or your own format.
- Replay: IQ keeps each answer in memory while it is being written. A browser that reconnects resumes from the exact event it missed. 0 events were lost on reconnect in the benchmark.
- Performance: 42,000 events per second on one CPU, 76,000 on two, and a p99 latency of 2.4 ms at 300 concurrent agent runs: more than 7× the throughput of general-purpose servers (Mercure, Laravel Reverb and Soketi) on the same CPU.
- Multi-tenant: separate credentials, limits and live counters for every app on the same servers.
- One Docker image for amd64 and arm64. No database, no Redis, no disk and no message broker.

## Pages

- [Home](https://iq-server.dev/): What IQ-Server is, how stream replay works, and how to integrate it.
- [Performance](https://iq-server.dev/performance/): Benchmark against Mercure, Laravel Reverb and Soketi: throughput, latency, scaling and memory.
- [For cloud providers](https://iq-server.dev/cloud-providers/): Offering LLM streaming as a managed, multi-tenant service, with a sizing and cost example.

## Contact

- [Talk to us](https://form.typeform.com/to/Ba6XOU1t): Ask about running IQ-Server, or about running the benchmark on your own infrastructure.