Atlas

Models

← All models

DiffusionGemma 26B A4B IT

Google · 2026-06-10 · 25.8B parameters

DiffusionGemma 26B A4B IT is Google DeepMind's instruction-tuned open-weight discrete-diffusion model built on a sparse Gemma 4 MoE foundation. It uses an autoregressive encoder and bidirectional diffusion decoder, supports configurable reasoning, and accepts interleaved text, image, and video input.

Benchmark scores

BenchmarkScore
AA-LCR14.3
AA-LCR v1.119.7
AA-Omniscience Index-59.2
Artificial Analysis Agentic Index2.2
Artificial Analysis Coding Index19.7
Artificial Analysis Intelligence Index13.5
Artificial Analysis Omniscience Accuracy16.6
Artificial Analysis Omniscience Hallucination Rate90.9
Artificial Analysis Openness Index44.4
Artificial Analysis Progress Index v1207
Capability131.7
CritPt (avg@5)0.3
GDPval-AA v2501
GDPval-AA v2 (normalized)2.7
GDPval-AA v2.1329
GPQA Diamond (Artificial Analysis avg@5)66.9
Humanity's Last Exam (Text-Only)10.8
IFBench (Artificial Analysis avg@5)59.5
MMMU Pro66.5
SciCode (Artificial Analysis avg@3)34.3
Terminal-Bench 2.112.4
𝜏³-Banking6.6
Loading Atlas data…