Bongard-mini
huggingface.co- Category
- AI Tools
- Rank
- No. 2602Tools index
Previous survey · No. 2487 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Models: Train & Run
- Date
About
Bongard-mini is an open-weight model built on the T5Gemma 2 4B-4B encoder-decoder architecture. It reads text, structured data or images plus questions with candidate answers or actions, and returns a probability distribution for each question without generating text. The model card shows it choosing actions in games such as 2048, Tetris, Othello, Snake and Doom.
At a glance
Bongard-mini scores answers you supply instead of writing its own. It is worth trying when several bounded decisions share one text or image and you can check its scores against known outcomes.
Good fit
Developers evaluating fixed-choice classification or ranking several proposed actions from the same evidence.
Consider something else when
Use a different approach if you need generated explanations or dependable verification of worked answers. The model card reports weak rejection of incorrect solutions.
Start here
In the documented playground, pick an example, replace the input with a fictional case and run two questions over it. Reorder the choices to see whether the scores move.
Cost and requirements
The hosted demo uses a limited anonymous ZeroGPU quota. The local weights are subject to Gemma terms, while the inference runtime uses Apache-2.0.
Limits to know
Scores can depend on choice order. Agreement with a benchmark teacher does not establish that a 70% score corresponds to a 70% chance of success in your application.
Sources checked 2026-11-05. Source-based assessment.
Tags
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.