strata-bench is an independent project. It is not part of Strata and is not maintained by its author.
Click a row for the full details: configuration, every request, charts and the raw file. Speeds are the engine's own timings, median of the measured runs (tokens per second, higher is better). The Medium, Long and XLong columns are prompt speeds for those prompts; decode speed is the medium prompt's. Results with a different configuration or engine are different measurements: compare rows that share the same model and configuration (the details page shows it). The first-token times and the agent-cache hit rate, which only standard-v2 results have, are on each result's details page and in the comparison; the two protocols are never ranked against each other on them. A badge such as lan means the client ran on another machine, so first-token times include the network.
Loading the reports...
The health view lists only catastrophic failures that the runner recorded: the server crashed, ran out of memory, returned degenerate output, hung, or answered with a 5xx error. A client-side problem or an interrupted run is never published.
No results yet
Be the first. Run the benchmark against your own Strata server and send the result file in a pull request; the guide takes about ten minutes.