Editorial Standards

ComputeFit publishes practical guidance on choosing and running local AI: hardware, models, quantization, benchmarks, and workflows. This page explains the standards behind that work.

Independence and commercial relationships

We do not accept payment from hardware vendors, model developers, or cloud providers in exchange for recommendations, rankings, or favorable test results. Commercial relationships — including any affiliate links we include in an article — never determine which products we recommend or how they are ranked, and where a link is present it is clearly labeled as such within the article.

On the use of AI tools

AI tools may assist with research, drafting, coding, or data processing behind our articles. All published conclusions are reviewed before publication, and editorial responsibility remains with ComputeFit.

Reproducibility

Where we report original measurements, we aim to include what a reader needs to replicate them: the model name and exact version (or file identifier), quantization level, hardware configuration, the software versions that matter for the result, and — where feasible — the full command or settings used. Where a number comes from another publication rather than our own machine, we say so directly in the article instead of letting it pass as an original measurement.

Dated claims

Model capabilities change quickly with every new release, and prices move just as fast. Figures that depend on a specific model build, runtime version, or market price are stated as belonging to that build or date rather than presented as permanent facts. When a retest after publication changes a number materially, we update the original article; when we have not yet been able to retest, we say so instead of implying the numbers are still current.

Plain language

We write for people whose main job is something other than machine learning research. If a term needs jargon, it gets defined on first use or linked to our own longer explanation; and precision that changes your decision — tokens per second, peak memory during a run, VRAM versus system RAM — gets reported rather than rounded away into a general impression of "runs fine."

Corrections

When we identify an error after publication, we correct the article and note the correction, rather than silently rewriting the record.