Open data
The llm-speed dataset
Every benchmark on this site is a real, cryptographically signed measurement. Here is the whole corpus in one place: 220 runs across 52 models and 10 hardware classes, free to download and reuse under CC BY 4.0. Each row carries the model, backend, workload, accelerator, decode tok/s, and a link to its signed run.
Last updated 2026-07-08 · suite suite-v1 · see the methodology for how each number is measured and signed.
Bulk download
The same corpus is mirrored as a HuggingFace dataset (with a live Dataset Viewer), for notebooks and training pipelines.
Prefer the live feed? The full JSON API returns every run, and per-run detail is at api.llm-speed.com/v1/results/{id}. Every run also has a citable page at llm-speed.com/r/{id}.
Reuse it (attribution required)
The data is licensed Creative Commons Attribution 4.0. Use it anywhere, including commercially, as long as you credit llm-speed with a link back. Copy one of these:
Data from llm-speed (https://llm-speed.com/data), CC BY 4.0.<a href="https://llm-speed.com/data">Benchmark data by llm-speed</a>, CC BY 4.0.Building a tool on this data (a VRAM calculator, a price-per-token comparison, a hardware guide)? That is exactly what the license is for. A link back is all we ask.
Citing it in a post or paper? Use:
llm-speed. The llm-speed dataset: signed LLM inference-speed benchmarks. https://llm-speed.com/data (2026). CC BY 4.0.Why the numbers are trustworthy
Most published tok/s figures are screenshots with no way to check them. Here every run is measured under the same suite-v1 workload, fingerprinted to real hardware, and cryptographically signed on upload, so a number cannot be quietly edited after the fact. If two rows disagree for the same model and GPU, both are real submissions from different setups, and you can open each permalink to see the full per-workload detail behind the headline decode rate. That is the difference between crowdsourced folklore and a dataset you can build on.
What is in each row
idandpermalink:the signed run and its citable pagetop_model_name,top_backend,top_workload:what was runaccelerator_summary:the GPU or Apple Silicon plus hosttop_decode_tps:the headline decode tokens per secondreceived_at,suite_version:when it landed and under which suite
The numbers are measured, not modeled. If a run looks wrong, open its permalink and check the raw workload results. Browse the same data as a leaderboard on the cheatsheet.