Instructions to use google/gemma-4-E2B-it-qat-mobile-ct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use google/gemma-4-E2B-it-qat-mobile-ct with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("google/gemma-4-E2B-it-qat-mobile-ct") model = AutoModelForMultimodalLM.from_pretrained("google/gemma-4-E2B-it-qat-mobile-ct", device_map="auto") - Notebooks
- Google Colab
- Kaggle
How to reproduce Gemma 4 mobile QAT .litertlm from public checkpoints?
Hi, thanks for the Gemma 4 mobile QAT release.
I'm trying to understand the intended public path for producing the official mobile .litertlm format from the released Gemma 4 mobile QAT checkpoints.
Specifically, I’m interested in the low-memory mobile artifact described for LiteRT-LM: the mixed-precision layout with targeted 2-bit weights and int8 activations, used for the <1GB runtime memory profile.
Could you confirm:
Is the official mobile .litertlm artifact reproducible from the public Gemma 4 mobile QAT checkpoints using public tooling?
If yes, what is the recommended checkpoint, command, and quantization recipe?
If not, is the .litertlm build currently produced by an internal Google pipeline, with the prebuilt litert-community/... .litertlm model being the intended artifact for deployment?
For fine-tuned Gemma 4 E2B models, what is the recommended public conversion recipe today to get closest to the official mobile memory footprint?
Thanks!
Same question...