Gemma 4 31B QAT
Google · released Jun 5, 2026
| size | 30.7 B |
| context window | 256,000 tokens |
| longest response | not stated by the maker |
| knowledge cutoff | January 2025 |
| open weights | yes, license: Apache 2.0 |
| how it ran | run locally on my computer |
| input and output | text and images in, text out |
Manufacturer data (model card, announcement), verified Sep 25, 2026.
Where we tested it
The train test
TaskThe model had to play a scene with a courier on a night train and go through a fixed script of my ten turns.
About 5 points out of 8, more than Claude Haiku 4.5. It caught the trap in turn five, but in turn eight it gave in and, on its own, invented a past in which the claim I slipped in holds true.
You'll find the size band and figures for other models on the Tested models page.