gemma-4-E2B-it-qat-q4_0-unquantized
Gemma 4 E2B instruction-tuned model, trained with Q4_0 quantization-aware training (QAT), mirrored as unquantized half-precision safetensors.
What it is for
This repo is a mirror of Google DeepMind's Gemma 4 E2B instruction-tuned QAT Q4_0 checkpoint. QAT trains the model to keep its quality when its weights are stored at Q4_0 precision. It holds the half-precision weights from the QAT pipeline in safetensors form, with the tokenizer, chat template, processor config and generation config. Use it as a pinned copy of the upstream weights for local inference and for further conversion.
Provenance
- Upstream model: google/gemma-4-E2B-it.
- Upstream QAT release: Gemma 4 QAT Q4_0 collection by Google DeepMind.
- Technical report: Gemma 4 Technical Report.
Use
Load it with Transformers.
from transformers import AutoProcessor, AutoModelForMultimodalLM
MODEL_ID = "chibifire/gemma-4-E2B-it-qat-q4_0-unquantized"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(MODEL_ID, dtype="auto", device_map="auto")
For prompts, chat template, thinking modes and multimodal input, read the upstream model card and the Gemma documentation.
Licence
Apache 2.0, under the Gemma 4 licence terms at https://ai.google.dev/gemma/docs/gemma_4_license. These terms travel with the weights. Keep this link when you copy or redistribute them.