Intel Google Gemma 4 E2B IT INT4 OpenVINO

This repository contains a Gemma 4 E2B Instruct OpenVINO GenAI INT4 model package for local edge inference on Intel hardware.

Model Details

Files

This is not a single-XML OpenVINO model. It is an OpenVINO GenAI model folder. Load the whole folder with VLMPipeline.

Important files include:

openvino_language_model.xml
openvino_language_model.bin
openvino_tokenizer.xml
openvino_tokenizer.bin
openvino_detokenizer.xml
openvino_detokenizer.bin
openvino_vision_embeddings_model.xml
openvino_vision_embeddings_model.bin
openvino_text_embeddings_model.xml
openvino_text_embeddings_model.bin
config.json
processor_config.json
tokenizer.json
tokenizer_config.json
chat_template.jinja

Usage

import openvino_genai as genai

model_dir = "intel_google-gemma-4-e2b-it-int4"
pipe = genai.VLMPipeline(model_dir, "GPU")

result = pipe.generate(
    "Answer in one short sentence: what is OpenVINO?",
    max_new_tokens=64,
)
print(result)

Local Test Results

Tested on an Intel Core Ultra platform with OpenVINO 2026.3.0.

Device Result
CPU Passed
Intel GPU Passed
Intel NPU Failed during NPU compilation

The NPU failure observed locally was a compiler error related to duplicated names during compilation. CPU and Intel GPU are recommended for this package.

Notes

The original local conversion attempt from the base checkpoint did not complete on the test machine because of memory pressure during export. This package was prepared from the pre-converted OpenVINO INT4 source package listed above and validated locally with OpenVINO GenAI.

License

Please review the license and terms of the base Gemma model and the source OpenVINO package before use or redistribution.

Downloads last month
1
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Advantech-EIOT/intel_google-gemma-4-e2b-it-int4

Finetuned
(208)
this model