Kapteeni v1 (meticulous)
Kapteeni, by Zetian Lin (TriusAI): a LoRA-tuned Qwen3-4B with small readout heads that serves the Jev System One wire format. Each question, option or score level gets its own pass; answers are probabilities, never generated text. Weights CC BY-SA 4.0.
Built by one person, Zetian Lin, under the TriusAI label. Small MLP readout heads sit on the last token's hidden state, with temperature scaling per question type. Siblings: kapteeni-v1-intuit (more accurate on well-formed numeric, temporal and multi-step traffic, less well calibrated) and kapteeni-v1.1c (2 October, on Qwen3.5-4B, one attached image, English and Chinese). On the third-party JevBench v1.4 public half, the maker's own runs give 0.710, 0.714 and 0.766 for the three, which the maker says are not an official rank. English-first; weak on long policies, multi-hop and temporal or numeric questions. The weights are CC BY-SA 4.0 because of share-alike training data (MultiNLI, FEVER); the code is Apache-2.0.
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
At a glance
| Parameters | 4B |
| Base model | Qwen/Qwen3-4B-Instruct-2507 |
| Maker | TriusAI |
| Released | 2026-09-25 |
| License | cc-by-sa-4.0 |
| Reported accuracy | 71.0% |
Get the weights
pip install systemonemodels
systemone pull triusai/kapteeni
The files are served from the maker's Hugging Face repository, TriusAI/kapteeni-v1-meticulous, and verified against the checksums recorded here.
Read more
This page was opened by System One for TriusAI, who can claim the organisation and take it over at any time.