Atlas

Models

← All models

Gemma 3 27B PT

Google · 2025-03-12 · 27.4B parameters

Gemma 3 27B PT is Google's open-weight, pre-trained multimodal checkpoint in the Gemma 3 family, released on March 12, 2025 alongside its instruction-tuned variant. Its 27,432,406,640-parameter checkpoint accepts text and images and generates text with a 131,072-token context window, and was trained on 14 trillion mixed image-and-text tokens using knowledge distillation. Google reports an August 2024 knowledge cutoff.

Benchmark scores

BenchmarkScore
AGIEval English66.2
ARC-Challenge70.6
ARC-Easy89.0
BBH (BIG-Bench Hard)77.7
BoolQ82.4
Capability113.4
DROP77.2
Global MMLU Lite — Google Unspecified Language Cohort75.7
GPQA Diamond24.3
GSM8K82.6
HellaSwag (Unspecified Scoring Protocol)85.6
HumanEval48.8
MATH50.0
MBPP65.6
MGSM74.3
MMLU78.6
MMLU Pro52.2
MMMU56.1
Natural Questions (open-domain, closed-book)36.1
PIQA83.3
RealWorldQA - release-reported visual perception53.9
Social IQa (SIQA)54.9
TriviaQA85.5
WinoGrande78.8
Loading Atlas data…