OnDevice.AI now has its first complete shortlist. The goal from the beginning was not to collect every small model or simply repeat the top row of a leaderboard. It was to answer a narrower and more useful question: which open or open-weight models are genuinely worth trying on a modern phone with 8 GB of RAM? After several rounds of model, licensing, download, and presentation checks, the first baseline is ready to be treated as a release rather than a draft.
The release covers eight jobs: text generation, coding, reasoning, image generation, speech-to-text, text-to-speech, photo analysis, and the best overall on-device model. Qwen3.5 4B handles the text-heavy categories, while Gemma 4 E4B Instruct is the broader multimodal and overall recommendation. Whisper Large V3 Turbo and Kokoro 82M cover speech, and LCM Dreamshaper v7 provides a permissively licensed image generation route that remains plausible within the device constraint.
What complete means
Complete does not mean permanent. Small-model capabilities, mobile runtimes, and quantization support move quickly. It means the selection rules and publishing structure are now stable enough to support future updates. Every category has a named winner, an original source, a practical mobile format, a license check, and a concise explanation of why it fits. Q4_K_M remains the default recommendation for compatible language and multimodal models, while speech and diffusion models use the formats that actually suit their runtimes.
Licensing is part of that definition. The shortlist favors Apache-2.0 and MIT choices so it remains useful to independent developers, commercial prototypes, and public tools. That standard is why LCM Dreamshaper v7 replaced the initially stronger-looking SD Turbo option. Raw performance matters, but a model that cannot be used for the intended project is not the best practical recommendation.
A website, an API, and a record
The human-readable shortlist lives on the homepage, where the cards are deliberately compact. The complete structured record is available from the public JSON endpoint at ondevice.ai/api. It includes the evidence, exact model identifiers, original repositories, mobile downloads, companion files, and runtime notes that would make the homepage too dense. This release is also frozen in the archive so later changes can be compared with the first baseline.
The website source is public at github.com/jasonas/OnDeviceAI. Opening the code matters for the same reason as publishing the API: this should be a resource other people can inspect, reuse, question, and improve. The site remains intentionally simple, using only HTML, CSS, and vanilla JavaScript, with Apache rules for clean URLs and the static JSON API. There is no private application layer hiding how the recommendations reach the page.
From here, meaningful recommendation changes will receive a dated API snapshot, an archived page, and a short post explaining what moved and why. Some days may bring more than one review, but updates will only become releases when the evidence changes the practical answer. That creates a history that is useful rather than merely busy.
Version 2026-06-28 is the starting line: a small, inspectable set of models chosen for real on-device constraints, with enough detail to act on and enough transparency to challenge. The shortlist will evolve, but the standard stays the same. A recommendation should be capable, downloadable, legally usable, and comfortable enough to belong on the device it claims to serve.