Four kinds of evidence
Claims should link to the manufacturer or project documentation, identify the exact configuration, and give a date when availability or pricing matters.
We show inputs, units, arithmetic, and exclusions. A memory estimate is not a runtime benchmark or a guarantee that a model fits.
A benchmark needs original results, hardware and software versions, model and settings, a method, and limitations.
Editorial advice should explain the tradeoff, the alternative, and when keeping existing hardware makes more sense.
What a useful benchmark includes
- Exact CPU, GPU, memory, storage, operating system, and driver.
- Runtime version, model, quantization, context length, input, and relevant settings.
- Repeated runs for timing claims, with loading time separated where appropriate.
- Memory pressure, failures, and output quality alongside speed.
- Original logs, screenshots, or data sufficient to understand the result.
Published manufacturer results remain manufacturer results. Planned tests remain plans. Neither becomes a TokenByte measurement because it appears in our article. See the benchmark status and method.
Images are not test evidence
We distinguish product photography, screenshots, diagrams, and editorial illustrations. An illustration communicates the subject; it does not prove a device configuration or a hands-on test.
Corrections are part of the work
Send a correction with the article URL and supporting evidence to support@tokenbyte.co. Substantive corrections should be identified in the article. The original publication date stays visible when a guide is refreshed.
Buying links and independence
A commercial link must not replace an explanation of the purchase. Recommendations need drawbacks and alternatives. Read our affiliate disclosure for how those links are handled.