The first OnDevice.AI homepage answered the important questions, but it asked visitors to absorb too much before they could make a decision. Every model card included the model ID, maker, size, runtime, quantization, mobile-fit guidance, benchmark evidence, download links, companion files, and runtime notes. That information was accurate and useful, yet repeating the full package across eight categories made the page feel more like a technical inventory than a shortlist. On a phone, the accumulated detail turned a quick comparison into a long scroll.
This update starts with a simpler principle: one obvious decision per category. Each recommendation now leads with the category and winning model, followed by one sentence explaining the choice. The next line keeps only the specifications that matter most when judging whether a model can run locally: size, runtime format, and license. Original and mobile-format download links remain directly available. Everything else has been removed from the immediate scan path.
Less on the page, not less information
Simplifying the homepage does not mean throwing away the research. The public JSON API remains the complete structured record, including model identifiers, evidence, mobile-fit notes, exact runtime files, companion files, and deployment guidance. The blog remains the place for the more human explanation: why a model won, what tradeoffs were accepted, and why a recommendation changed. That division gives each format a clearer job. The homepage helps a person choose; the API helps software consume the resource; the blog explains the judgment behind it.
The page structure changed with that same goal in mind. The familiar card layout remains, but each card is smaller, quieter, and focused on a single recommendation. Repeated labels and secondary details have been reduced so model names and category differences carry the visual hierarchy. The opening copy is shorter, update metadata is more compact, and the latest decision note now appears under a clearer "How We Choose" heading. Update history and sources still sit at the bottom, where they remain easy to find without competing with the recommendations.
A navigation that respects a small screen
Mobile navigation was another source of avoidable weight. Four full-width navigation links previously occupied a grid above the headline, so the main content began only after a block of controls. The new header keeps the OnDevice.AI name, theme switch, and a familiar menu icon on one line. The links appear only when requested. The menu is keyboard accessible, reports its open state to assistive technology, closes after a link is selected, and still leaves the navigation visible if JavaScript is unavailable.
We considered expandable detail inside every model row, but that would still make the homepage responsible for two different jobs. It would also introduce eight more controls and encourage visitors to open several sections before comparing them. Sending deeper questions to a stable API and focused blog posts is calmer and more useful. A person who only needs the best coding model can get the answer quickly; a developer evaluating licenses and runtime files can follow the precise links without losing any information.
Importantly, this began as a presentation update, not a model change. The eight recommendations, permissive-license requirement, 8 GB target, and API schema stayed the same. The shortlist was then marked complete later on June 28. Keeping those concepts separate matters because OnDevice.AI may refresh several times in a day. A clearer, more stable reading pattern will make future model changes easier to notice, and the history will continue to show what changed in the data rather than treating every visual adjustment as a new model snapshot.
The result should take less effort to understand, especially on mobile. That is the standard we will keep using: the shortlist should feel short, while the evidence behind it stays complete and available to anyone who wants to look closer.