File size: 1,808 Bytes
f3e82a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d7d99ec
 
f3e82a2
d7d99ec
f3e82a2
d7d99ec
 
f3e82a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
---

license: gemma
base_model: google/gemma-4-E4B-it-qat-q4_0-unquantized
base_model_relation: quantized
tags:
- openvino
- int4
- qat
- intel
- arc
- igpu
- core-ultra
pipeline_tag: image-text-to-text
---


# Gemma 4 E4B QAT — OpenVINO INT4 (q4_0-matched scheme)



OpenVINO IR conversion of

[google/gemma-4-E4B-it-qat-q4_0-unquantized](https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized)

— Google's **quantization-aware-trained** Gemma 4 E4B checkpoint, converted with the

QAT-matched scheme (symmetric, group size 32):



```bash

optimum-cli export openvino -m google/gemma-4-E4B-it-qat-q4_0-unquantized \
  --task image-text-to-text --weight-format int4 --sym --group-size 32 \
  gemma-4-E4B-it-qat-int4-ov
```



QAT means this int4 build is trained to be quantized, so the quality loss is far smaller than

post-training quantization at the same bit width.



## Conversion note



Loader warnings about "missing" k/v projections on upper layers are Gemma 4's tied KV-shared

weights (verified benign). Serving and tooling: [core-ultra-llm-server](https://github.com/harmenwessels/core-ultra-llm-server).



## Usage (OpenVINO GenAI)



```python

import openvino_genai as ov_genai



pipe = ov_genai.VLMPipeline("gemma-4-E4B-it-qat-int4-ov", "GPU", CACHE_DIR="./.ovcache")

print(pipe.generate("Explain Python decorators in three sentences.", max_new_tokens=128))

```

VLM-shaped IR — requires `VLMPipeline` even for text-only use.

## Provenance & license

- Base: Google's Gemma 4 E4B QAT checkpoint (released 2026-04); weights are governed by the
  [Gemma Terms of Use](https://ai.google.dev/gemma/terms)
- Conversion date: 2026-06-06; optimum-intel git-master, transformers 5.5.0
- No finetuning — direct quantization of Google's QAT weights