OnDevice.AI starts from a simple question: what would I actually want to run locally on a modern phone with 8 GB of RAM? That question rules out a lot of impressive models immediately. A leaderboard can tell us what is strongest in the abstract, but it cannot tell us whether a model will feel sane inside a private, offline, battery-powered app. For this first shortlist, the selection is deliberately practical. We favored models that are small enough to deploy, have clear download paths, and can be used without sending personal data to a server.

The text-heavy categories converged quickly. Qwen3.5 4B is a strong fit for text generation, coding, and reasoning because it lives near the right size boundary for 8 GB devices while still posting serious small-model results. The Q4_K_M GGUF build is the kind of compromise that matters on phones: quality remains useful, the file is not absurd, and existing llama.cpp-style runtimes have a path to run it. It is not the biggest Qwen model, but that is the point. A model that never comfortably loads on the target device is not a good on-device recommendation.

The speech choices follow the same logic. Whisper Large V3 Turbo is not tiny, but it has the right mix of recognition quality, language coverage, permissive MIT licensing, and mature mobile runtime support through whisper.cpp and Core ML style deployment. For text-to-speech, Kokoro 82M is almost the opposite kind of pick: very small, Apache-2.0, and practical enough that it can sit beside a local assistant without consuming the whole memory budget.

Multimodal selection is trickier. Gemma 4 E4B remains the best overall and photo-analysis pick in this shortlist because it is explicitly aimed at mobile and edge use while covering text, image, and short audio input. Qwen3.5 4B is excellent for text, coding, and reasoning, and it has multimodal abilities too, but Gemma's broader edge-oriented coverage makes it a cleaner "one model for several private experiences" answer.

The licensing correction

The most useful change in this update is the image-generation pick. The first pass selected SD Turbo because it is fast and practical for mobile-class generation. The problem is the license: Stability AI's non-commercial community license is not a good fit for a resource that should be useful to developers, commercial experiments, and public tools. A recommendation is not just a quality statement. It is also a deployment statement.

That pushed the image category toward LCM Dreamshaper v7. It is MIT-licensed, available as a Diffusers text-to-image model, and the repository includes ONNX assets for a more mobile-friendly deployment path. Is it the absolute best image model? No. FLUX.1-schnell is Apache-2.0 and stronger on quality, but it is far too large to call comfortable on an 8 GB phone. SD Turbo is also likely ahead of LCM Dreamshaper v7 in prompt following and visual fidelity. But SD Turbo brings a non-commercial license, and FLUX brings a size mismatch. LCM Dreamshaper v7 is the best fit for the actual constraint: permissive license, fast generation, and realistic mobile deployment.

What the shortlist is really optimizing for

The ranking is not a pure benchmark table. It is a filtered set of recommendations for local, private AI. The filters are:

Those filters sometimes disagree. The highest-quality model may be too large. The fastest model may carry an awkward license. The most famous checkpoint may lack a clean mobile runtime. The goal of OnDevice.AI is to make those tradeoffs visible, not hide them behind a single "best" label.

This work became the first complete shortlist on June 28, 2026. Future meaningful updates will get their own short posts too, so the site has a readable record of what changed, why it changed, and what tradeoff moved the recommendation.