Bekko System One v0
Ultra-small English encoders (17M, 68M and 395M) fine-tuned from Ettin rerankers. Each answers choice, noul and score questions with a probability per candidate and no generated text. Experimental v0 by one developer; the weights have no licence assigned yet.
Full fine-tunes of Ettin rerankers with a shared-prefix encoder: state and instructions are encoded once, every candidate attends to that prefix, and choice, noul and score heads give a softmax over the supplied candidates. English only; probabilities are not calibrated. The maker's own S1MB benchmark (a baseline-adjusted task average, not accuracy) gives 50.60 for the 400M against 59.59 for Jev 1.13; the maker says training shares 77 subset names with S1MB and that the models trail Jev on generalisation, hence "v0". The third-party Decision Index 0.3 puts the 400M at 15.24. Siblings bekko-system-one-v0-17m and -68m, with ONNX browser exports. No licence is assigned to the weights yet.
What it decides
- choice — picks one option from a set
- noul — answers a yes/no question with one probability
- score — places the input on an ordered scale
At a glance
| Parameters | 395M |
| Base model | cross-encoder/ettin-reranker-400m-v1 |
| Maker | Yuichi Tateno (hotchpotch) |
| Released | 2026-09-29 |
| License | Unassigned: the weights licence is not finalised (treated as non-commercial); training code MIT |
| Reported latency | 5.03 ms median per question on an RTX 5090 |
Get the weights
pip install systemonemodels
systemone pull hotchpotch/bekko-system-one
The files are served from the maker's Hugging Face repository, hotchpotch/bekko-system-one-v0-400m, and verified against the checksums recorded here.
Read more
This page was opened by System One for Yuichi Tateno (hotchpotch), who can claim the organisation and take it over at any time.