Atlas

Models

← All models

Gemma 3n E2B Instruct

Google DeepMind · 2025-06-26 · 5.4B parameters

Gemma 3n E2B IT is the smaller standalone instruction-tuned checkpoint from Google's mobile-first Gemma 3n family, fully released on June 26, 2025. Its 5.44B-parameter checkpoint combines a 30-layer MatFormer text decoder with Per-Layer Embeddings, LAuReL, and AltUp plus MobileNet-V5 vision and USM-derived audio encoders; it accepts text, image, audio, and video in a 32K-token context and generates text. Google reports that PLE caching and parameter skipping can reduce its effective memory load to a rounded 1.91B parameters, enabling operation in as little as 2 GB, with text training across more than 140 languages and multimodal understanding across 35 languages.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-79.6
AIME 202510.3
Artificial Analysis Intelligence Index1.0
Artificial Analysis Omniscience Accuracy6.7
Artificial Analysis Omniscience Hallucination Rate92.4
Artificial Analysis Openness Index50.0
Capability103.8
CritPt0.0
GPQA Diamond22.9
Humanity's Last Exam (Text-Only)4.0
IFBench (Artificial Analysis)22.0
LiveCodeBench (Artificial Analysis)9.5
MATH-500 (Artificial Analysis source)69.1
SciCode (Artificial Analysis)5.2
Terminal-Bench Hard0.8
τ²-bench (Telecom)0.0
Loading Atlas data…