Bongard-mini
Open-weight judgment model on T5Gemma 2 4B-4B: reads text, records or images once and returns a probability for every candidate answer to noul, choice and score questions, without generating text. Weights under the Gemma Terms of Use.
The encoder reads the state once and decoder branches answer up to eight questions in parallel through a trained judgment head; its own /v1/systemone server reports zero output tokens. Choice takes 2 to 255 candidates, score 2 to 10 levels. Choice probabilities can depend on option order (a rotations setting averages over cyclic orders); use the shipped temperatures and refit calibration for a new domain. The full model needs the bongard runtime; the backbone alone in Transformers is not the judgment model. On JevBench's public items (a third-party benchmark) AgentBull's own run gives 100% easy, 97.2% standard and 57.7% hard; the card cites 4th of 61 on the third-party Hanno Labs DecisionBench leaderboard (29 September snapshot). Technical report by Li Ding, Haidi Jin and Chen Ji. Weights are Gemma derivatives: commercial use is allowed under the Gemma Terms, and any mirror must carry those terms and the NOTICE.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 7.5B |
| Base model | google/t5gemma-2-4b-4b |
| Maker | AgentBull |
| Released | 2026-09-27 |
| License | Gemma Terms of Use (weights, incl. Prohibited Use Policy); runtime code Apache-2.0 |
| Reported accuracy | 78.0% |
| Reported latency | 36.1 ms median per decision on short JevBench requests, one RTX PRO 6000, BF16 |
Get the weights
pip install systemonemodels
systemone pull agentbull/bongard-mini
The files are served from the maker's Hugging Face repository, AgentBull/bongard-mini, and verified against the checksums recorded here.