DiffusionGemma
An experimental open model that explores an exceptionally fast approach to text generation
DiffusionGemma is an experimental open-weights generative model family built by Google DeepMind. It is based on the 26B A4B Mixture-of-Experts Gemma 4 architecture and uses discrete text diffusion / block-autoregressive multi-canvas sampling to generate text faster while accepting multimodal text, image, and video inputs.
Pricing
Model profile · Current snapshot
Input / 1M
$0.02
Output / 1M
$0.10
Blended / 1M
$0.04
Output TPS
82.89
TTFT (s)
0.35
Model Intelligence
Context window
256,000 tokens
Arena ranking
12
Benchmarkable
No
Model level
family
Intelligence Index
12.2
Coding Index
9.4
GPQA
0.58
HLE
0.04
SciCode
0.24
IFBench
0.44
LCR
0.31
TerminalBench Hard
0.08
TAU2
0.21
Recent stories
0 linked stories
No linked stories yet.