Saracura PT-BR 4B
A Brazilian Portuguese decision model on the Kev recipe: a LoRA adapter and pointer head on Qwen3.5-4B-Base, initialised from kev-4b and fine-tuned on Portuguese decisions. Answers choice, yes/no and score questions in one pass and serves /v1/systemone.
The test split comes from the same sources as training; the only task outside training (a Banco Central FAQ) scores 96.8%. Claim checking on FACTCK.BR is near chance, and the maker says confidence is not calibrated on the user's data. Needs a GPU, or Apple Silicon through MLX. head.pt is a PyTorch pickle. The maker's own run of the third-party Decision Index gives 37.91, submitted but not yet accepted. Sibling saracura-ptbr-v0 (322M, on Laya multilingual).
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
At a glance
| Parameters | 4B |
| Base model | jaredpalmer/kev-4b |
| Maker | Felhen |
| Released | 2026-10-06 |
| License | apache-2.0 |
| Reported accuracy | 73.1% |
| Reported latency | 100 to 170 ms per decision on an RTX 5090 in bf16 |
Get the weights
pip install systemonemodels
systemone pull felhen-ai/saracura-ptbr
The files are served from the maker's Hugging Face repository, felhen-ai/saracura-ptbr-4b, and verified against the checksums recorded here.
Read more
This page was opened by System One for Felhen, who can claim the organisation and take it over at any time.