See our collection for all versions of Gemma 3.

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

kerasformers/gemma-3-270m-it

Paper: Gemma 3 Technical Report (arXiv:2503.19786) · HF Papers

Gemma 3 is Google's open Gemma family with a text-only decoder: RMSNorm, GeGLU, RoPE with a long-context rope scaling, grouped-query attention, alternating local/global attention, and (4B+) a SigLIP image encoder feeding soft image tokens into the decoder. Base checkpoints are for completion; -it variants are instruction-tuned.

For more details, see Google's original model card.

Pure-Keras 3 conversion of google/gemma-3-270m-it for kerasformers. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an instruction-tuned checkpoint: use the chat template via Gemma3Tokenizer.

Quick start

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

from kerasformers.models.gemma3 import Gemma3Generate, Gemma3Tokenizer

model = Gemma3Generate.from_weights("kerasformers/gemma-3-270m-it")
tokenizer = Gemma3Tokenizer.from_weights("kerasformers/gemma-3-270m-it")

inputs = tokenizer([
    {"role": "user", "content": "Explain rotary embeddings in one sentence."}
])
outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0]))

All Gemma 3 variants load the same way with from_weights("kerasformers/<variant>"):

Variant	Hub	Type
`gemma-3-270m`	`kerasformers/gemma-3-270m`	text / base
`gemma-3-270m-it`	`kerasformers/gemma-3-270m-it`	text / instruct
`gemma-3-1b-pt`	`kerasformers/gemma-3-1b-pt`	text / base
`gemma-3-1b-it`	`kerasformers/gemma-3-1b-it`	text / instruct
`gemma-3-4b-pt`	`kerasformers/gemma-3-4b-pt`	multimodal / base
`gemma-3-4b-it`	`kerasformers/gemma-3-4b-it`	multimodal / instruct
`gemma-3-12b-pt`	`kerasformers/gemma-3-12b-pt`	multimodal / base
`gemma-3-12b-it`	`kerasformers/gemma-3-12b-it`	multimodal / instruct
`gemma-3-27b-pt`	`kerasformers/gemma-3-27b-pt`	multimodal / base
`gemma-3-27b-it`	`kerasformers/gemma-3-27b-it`	multimodal / instruct

Tips

Set KERAS_BACKEND before importing Keras / kerasformers.
Loads in bfloat16 by default (the weights are bf16). Pass load_dtype="float32" for full precision, or quantization="int8" to shrink further.
gemma-3-1b is text-only — use gemma-3-4b-it or larger for images.
See Gemma 3 docs and Loading Weights.
Upstream safetensors still work via the hf: prefix, e.g. Gemma3Generate.from_weights("hf:google/gemma-3-270m-it").