Benchmarks you can actually understand
No original benchmark results have been published yet. This page sets out how we intend to measure performance so that a result has useful context.
Record the whole setup
Publish the exact hardware model, chip, memory, storage, operating system, tool version, power mode, and whether the machine is connected to power. Note background applications and thermal conditions.
Choose a real, repeatable workload
For development, record a fixed project revision and separate clean builds from incremental builds. For local AI, record the model, quantization, prompt, context size, and runtime settings. Keep model-loading time separate from generation time.
Repeat and show the variation
Run a documented warm-up followed by at least five measured runs. Publish each measurement, the median, and the range. Explain failed runs and exclusions rather than quietly removing inconvenient results.
Publish the recipe
Include commands, project or model references, raw data, and limitations. Compare like with like. A synthetic score, a real project build, and an AI response rate answer different questions.
The illustration on the homepage is a checklist graphic, not measured performance data. Hardware access and original testing are still needed before comparisons can be published.