Atlas

Models

← All models

Gemma 4 E4B IT

Google DeepMind · 2026-04-02 · 8B parameters

Gemma 4 E4B IT is Google DeepMind's instruction-tuned, dense, edge-oriented Gemma 4 checkpoint, released on April 2, 2026 under Apache 2.0; the official checkpoint repository reports 7,996,156,490 total parameters. Its per-layer embeddings yield an approximately 4.5-billion-parameter effective footprint rather than a sparse active-parameter count; it accepts text, images, audio, and video frame sequences, generates text across a 131,072-token context window, and offers caller-controlled thinking. Google states the pretraining-data cutoff only as January 2025, so the structured knowledge cutoff remains null.

Benchmark scores

BenchmarkScore
AA-LCR30.7
AA-Omniscience Index-20.1
Artificial Analysis Agentic Index1.8
Artificial Analysis Coding Index9.4
Artificial Analysis Intelligence Index12.2
Artificial Analysis Omniscience Accuracy8.6
Artificial Analysis Omniscience Hallucination Rate31.3
Artificial Analysis Openness Index38.9
Capability131.1
CritPt0.6
GDPval-AA v2230
GDPval-AA v2 (normalized)0.0
GPQA Diamond57.6
Humanity's Last Exam (Text-Only)4.7
IFBench (Artificial Analysis)44.2
MMMU Pro51.4
SciCode (Artificial Analysis)24.4
Terminal-Bench 2.11.9
Terminal-Bench Hard8.3
τ²-bench (Telecom)26.0
𝜏³-Banking5.4
Loading Atlas data…