Modelpublic

tuanhdev/vidroidcall-qwen3-0.6B-nlu-gguf-v6

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

1likes254downloads

Model Card

ViDroidCall NLU — Qwen3 0.6B (Q4KM)

On-device Vietnamese NLU for [ViDroidCall Studio](https://github.com/tuanhdevvn/ViDroidCall-Studio) (Apache-2.0).

Fine-tune of Qwen3-0.6B, exported as GGUF Q4_K_M.

—App source: https://github.com/tuanhdevvn/ViDroidCall-Studio
—This repo stores every shipped train version. Do not delete older files — docs and older app builds pin them by filename.

Recommended file: qwen3-nlu-run-014-Q4_K_M.gguf

GGUF is weights only. The Android app must inject the NLU system prompt and run the validator after inference (temperature=0).

Versions

File	Run	Intents	Size	Note
`qwen3-nlu-run-014-Q4_K_M.gguf`	run-014	11 (includes `search_web`)	~378 MB	Latest. MVP pass, hard-v3 93.5%
`qwen3-nlu-run-006-Q4_K_M.gguf`	run-006	10	~397 MB	Demo / contest release. Keep.

When a new train is ready, add qwen3-nlu-run-XXX-Q4_K_M.gguf to this repo and add a row here. Do not overwrite an existing filename.

Install on Android

Latest:

bash

adb push qwen3-nlu-run-014-Q4_K_M.gguf /sdcard/Download/

Contest / older docs (run-006):

bash

adb push qwen3-nlu-run-006-Q4_K_M.gguf /sdcard/Download/

The app scans .gguf in the device Download folder. Fast-Path still works without this file. If both files are in Download, keep only the version you want the app to load.

License

Apache-2.0. Base model: Qwen3 (Alibaba).