Atlas

Models

← All models

Gemma 4 26B A4B IT

Google DeepMind · 2026-04-02 · 26.5B parameters

Gemma 4 26B A4B IT is Google's instruction-tuned checkpoint of the 26B-class Gemma 4 sparse Mixture-of-Experts model, with 128 routed experts plus one shared expert and eight routed experts selected per token. It accepts interleaved text, image, and video input, supports configurable thinking and a 262,144-token context window, and generates text.

Benchmark scores

BenchmarkScore
AA-LCR55.7
AA-Omniscience Index-48.1
ALE-Bench1165
Artificial Analysis Agentic Index11.0
Artificial Analysis Coding Index39.3
Artificial Analysis Intelligence Index25.7
Artificial Analysis Omniscience Accuracy18.1
Artificial Analysis Omniscience Hallucination Rate80.9
Artificial Analysis Openness Index38.9
BRIDGE Medical (chain-of-thought)45.2
BRIDGE Medical (few-shot)53.2
BRIDGE Medical (zero-shot)44.5
BullshitBench v2 Leaderboard21.0
Capability145.5
Creative Writing v3 Elo1291
Creative Writing v3 Rubric Score16.0
CritPt0.0
EQ-Bench 31431
EQ-Bench 3 (Rubric Score)70.0
EQ-Bench 3: Analytic9.2
EQ-Bench 3: Assertive6.7
EQ-Bench 3: Compliant5.8
EQ-Bench 3: Empathy7.8
EQ-Bench 3: Humanlike5.3
EQ-Bench 3: Insight7.8
EQ-Bench 3: Moralising5.0
EQ-Bench 3: Pragmatic7.7
EQ-Bench 3: Safety6.6
EQ-Bench 3: Social IQ6.9
EQ-Bench 3: Warm6.9
Loading Atlas data…