The 45 repos behind github.com/john-rocky/windows-zoo: ONNX, OpenVINO IR and precompiled Intel NPU blob per model, measured on one Windows laptop.
Daisuke Majima (MLBoy) PRO
mlboydaisuke
AI & ML interests
On-device AI. Core AI model zoo (60+ models) · devicemark iPhone LLM leaderboard · author of "The Art of Core AI" → github.com/john-rocky
Recent Activity
posted an update about 4 hours ago
MNN-Pocket-0.3B-Dec on a Galaxy S26: your app can ask its own multiple-choice questions and get each option's probability, with no LLM API. The NPU kept every tested probability within 0.02 of my PyTorch run of the original model.
https://huggingface.co/litert-community/MNN-Pocket-0.3B-Dec-LiteRT
A model from Alibaba's MNN team for multiple-choice, yes/no and rating questions, converted to LiteRT.
In the clip, a test app answers questions in 13.5, 15.6 and 13.7 ms on the NPU, plus 9.5 to 17.6 ms in its own code.
To run it: the card's Python host; for Android, port it as the test app did. posted an update 1 day ago
MNN-Pocket-0.3B on a Galaxy S26: an assistant in your app can turn a short request into a team note in under 0.7 s, without a server. It decodes about 105 tokens/s on the GPU.
https://huggingface.co/litert-community/MNN-Pocket-0.3B
The MNN team's chat model for on-device assistants, data extraction and light agent pipelines (not heavy reasoning), converted to LiteRT-LM.
In the clip, the GPU writes the note with the request's day, time and place; a test app shows 674.0 ms from Send to the last token.
To run it: the card has LiteRT-LM command-line and Python commands. updated a collection 1 day ago
LiteRT Conversions - shipped to litert-community