Daily leaderboard checks continue at OnDevice.AI. On October 30, 2026, we took another deeper look at the eight categories, including the original model cards, licenses, and actual runtime downloads. This time the shortlist changes: Qwen3.5 4B replaces Gemma 4 E4B Instruct for photo analysis and best overall. The release also qualifies our image-generation recommendation and records several serious challengers for future comparisons.
Our standard remains offline, private use on a phone with 8 GB of RAM. Arena and Artificial Analysis provide useful quality signals, but their rankings do not certify mobile memory use or sustained speed. This update is a review of published evidence, not a new set of physical-phone benchmarks. That distinction matters when interpreting both the changes and the recommendations we retain.
The whole download matters
The Gemma package we previously linked contains a roughly 5.41 GB Q4_K_M file and a 990 MB multimodal projector. The Qwen package contains a roughly 2.74 GB Q4_K_M file and a 672 MB F16 projector. Together, those are about 6.40 GB and 3.41 GB of downloads respectively. The smaller package gives Qwen a more promising starting point for our memory budget.
File size is not peak RAM. The operating system, application, context cache, image processing, and inference buffers all need space, while loading and offloading strategies vary between runtimes. An effective parameter count also does not describe every weight a download contains. Our default advice now starts at 4K-8K context, with larger contexts earned through measurements on the intended phone. Long advertised context windows are capabilities, not mobile settings.
Why photo analysis moves to Qwen
Qwen3.5 4B accepts images as well as text. Its developer reports 66.3 on MMMU-Pro and 74.6 on MathVision, while Google reports 52.6 and 59.5 for Gemma E4B. These results support reconsidering our previous winner. They are separate developers' evaluations, however, and do not establish the quality of either phone quantization in a controlled comparison. The recommendation combines that quality signal with the smaller runtime package.
The same balance makes Qwen our overall default across text, coding, reasoning, and images. Gemma still offers a meaningful advantage for projects requiring native audio input; Qwen does not replace that capability. Voice applications can instead combine our overall pick with dedicated speech models, loading components as needed. The matching vision projector is only required when using images.
Coding has a credible challenger
Qwen remains our text, coding, and broad-reasoning pick, but coding is now a close review priority. MiniCPM5-2B has an Apache-2.0 license and an official Q4_K_M download around 1.56 GB. OpenBMB reports LiveCodeBench v6 scores of 69.1 for MiniCPM and 56.4 for Qwen in its comparison. Qwen leads on other measures in that table, including general knowledge and GPQA.
That evidence warrants a direct trial rather than an automatic swap across several categories. Useful mobile coding involves edit quality, instruction following, response time, and how much thinking output the task needs. We retain Qwen while recording MiniCPM as a leading comparison candidate. A higher server benchmark score alone does not tell us which quantized model will deliver the better phone experience.
Speech defaults stay, with alternatives
Whisper Large V3 Turbo stays for broad multilingual transcription and mature quantized deployment. For English live dictation, Moonshine Medium Streaming deserves comparison: it has 245M parameters and an MIT license. Kokoro 82M remains our lightweight speech-output default. Newer options such as Pocket TTS bring streaming, cloning, and very small October English checkpoints. Listening tests, language needs, and device latency should guide any eventual replacement.
A correction on image licensing
Our earlier description of LCM Dreamshaper v7 was too confident about a pure MIT license. Its repository is tagged MIT, but it explicitly identifies Dreamshaper v7 as its distillation source, and that upstream repository lists CreativeML OpenRAIL-M. The tag alone does not verify that the whole pipeline meets a strict MIT/Apache-only requirement. We now mark the image pick provisional and expose the qualification on its card and in the API. Projects with that requirement should wait for clarification or choose a separately verified pipeline.
Bonsai Image Ternary 4B is an Apache-2.0 candidate with compressed weights and published iPhone results. Those phone measurements used 12 GB of RAM, so they do not establish comfortable operation on our 8 GB target. We also correct our earlier blanket advice about image formats: stable-diffusion.cpp supports diffusion GGUF. A compatible diffusion runtime is essential; an LLM loader is not interchangeable with it.
The October snapshot and its JSON record preserve these decisions, alongside the older releases. Two winners change today, and one recommendation gains an important qualification. Daily checks continue, with future releases tied to evidence that improves the practical answer for the phone in your hand.