Atlas

Models

← All models

Gemma 3 12B PT

Google DeepMind · 2025-03-12 · 12.2B parameters

Gemma 3 12B PT is Google’s open-weight, pre-trained multimodal checkpoint in the Gemma 3 family, released on March 12, 2025. It accepts text and images and generates text, supports a 131,072-token context window and pretrained coverage of more than 140 languages, and was trained on 12 trillion mixed image-and-text tokens using knowledge distillation. Google reports an August 2024 knowledge cutoff.

Benchmark scores

BenchmarkScore
Antidote: Everyday Edition897
Global PIQA - Non-Parallel (Log-Likelihood Accuracy)64.2
Global PIQA - Parallel (Log-Likelihood Accuracy)29.3
Loading Atlas data…