See our collection for all Gemma 4 sizes and variants.

Run Gemma 4 with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/gemma-4-e2b-it

Pure-Keras 3 conversion of google/gemma-4-E2B-it for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX. This is the 5B variant, served here as image + audio + text -> text via Gemma4Processor; weights are stored in bfloat16.

For model details, license, and usage terms, see Google's model card.

Gemma 4 family

Property	E2B	E4B	12B Unified	31B Dense
Total Parameters	2.3B effective (5.1B with embeddings)	4.5B effective (8B with embeddings)	11.95B	30.7B
Layers	35	42	48	60
Sliding Window	512 tokens	512 tokens	1024 tokens	1024 tokens
Context Length	128K tokens	128K tokens	256K tokens	256K tokens
Vocabulary Size	262K	262K	262K	262K
Supported Modalities	Text, Image, Audio	Text, Image, Audio	Text, Image, Audio	Text, Image
Vision Encoder Parameters	~150M	~150M	-	~550M
Audio Encoder Parameters	~300M	~300M	-	No Audio

✨ Quick start

Text-only

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from zeromodels.models.gemma4 import Gemma4TextGenerate, Gemma4Tokenizer

model = Gemma4TextGenerate.from_weights("zeromodels/gemma-4-e2b-it")
tokenizer = Gemma4Tokenizer.from_weights("zeromodels/gemma-4-e2b-it")

inputs = tokenizer([{"role": "user", "content": "Hello, who are you?"}])
outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0]))

Image + audio + text

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.gemma4 import Gemma4ConditionalGenerate, Gemma4Processor

model = Gemma4ConditionalGenerate.from_weights("zeromodels/gemma-4-e2b-it")
processor = Gemma4Processor.from_weights("zeromodels/gemma-4-e2b-it")

inputs = processor(conversation=[
    {"role": "user", "content": [
        {"type": "image", "image": Image.open("cat.jpg")},
        {"type": "audio", "path": "clip.wav"},
        {"type": "text", "text": "Describe the image and what you hear."},
    ]}
])
outputs = model.generate(**inputs, max_new_tokens=64)
print(processor.decode(outputs[0]))

Load any Gemma 4 variant the same way with from_weights("zeromodels/<variant>"):

Variant	Hub
`gemma-4-12b`	zeromodels/gemma-4-12b
`gemma-4-12b-it`	zeromodels/gemma-4-12b-it
`gemma-4-26b-a4b`	zeromodels/gemma-4-26b-a4b
`gemma-4-26b-a4b-it`	zeromodels/gemma-4-26b-a4b-it
`gemma-4-31b`	zeromodels/gemma-4-31b
`gemma-4-31b-it`	zeromodels/gemma-4-31b-it
`gemma-4-e2b`	zeromodels/gemma-4-e2b
`gemma-4-e2b-it`	zeromodels/gemma-4-e2b-it
`gemma-4-e4b`	zeromodels/gemma-4-e4b
`gemma-4-e4b-it`	zeromodels/gemma-4-e4b-it

Tips

Set KERAS_BACKEND before importing Keras / zeromodels.
Loads in bfloat16 by default. Pass load_dtype="float32" for full precision, or quantization="int8" to shrink further.
See the Gemma 4 docs.
Community / upstream weights still work via the hf: prefix: Gemma4ConditionalGenerate.from_weights("hf:google/gemma-4-E2B-it").