# llamaperf

> Community speed reports for running LLMs on your own hardware. Each report is one person's measured tokens per second for one model, quantization and engine on one GPU or Mac, extracted into structured fields and linked to the post or write-up it came from.

llamaperf is not a lab benchmark. Speeds are what people reported. Every report records what kind of run it was (plain, speculative decoding, several concurrent requests, or part of the model offloaded to system RAM), and every median and "fastest" figure on the site counts plain runs only. A range in a source is kept as a range, never averaged into one number.

## Structured access

- Read the catalog with no key: MCP server at https://llamaperf.com/api/mcp/public (Streamable HTTP). Tools: find_hardware, find_models, search_reports, hardware_performance, check_fit.
- Measure a local model server with one file, no dependencies: https://llamaperf.com/llamaperf-bench.mjs (details at https://llamaperf.com/agents). It sends nothing unless asked.
- Submit a benchmark from your own machine: MCP server at https://llamaperf.com/api/mcp with an agent key or OAuth. Setup: https://llamaperf.com/submit/agent
- Every report page has a stable public URL of the form https://llamaperf.com/report/<uuid>.

## Pages

- [Local LLM, what it is and what to run](https://llamaperf.com/local-llm)
- [Best local LLM by memory size](https://llamaperf.com/best/local-llm)
- [Leaderboard by hardware tier](https://llamaperf.com/leaderboard)
- [VRAM calculator: what fits on your GPU or Mac](https://llamaperf.com/calculator)
- [Compare GPUs and models](https://llamaperf.com/compare)
- [GPU pages](https://llamaperf.com/gpu): speeds by model for one card
- [Model pages](https://llamaperf.com/model): speeds by card for one model family
- [Guides](https://llamaperf.com/guides/run-llm-locally)
- [Submit a benchmark](https://llamaperf.com/submit)

## Reading the numbers

- "Measured" is a community median at the same bit class. "Estimated" is the calculator's formula, and the two are labelled apart wherever both appear.
- A median with one or two reports behind it is one person's run, not a typical figure. Look at n.
- Reddit posts, GitHub and Hugging Face write-ups are the sources. A source link is on every report.