Manchego v3
Oraculum's 4B typed-decision model: a merged LoRA on Qwen3.5-4B that reads a state, a question and lettered options and returns a probability for each option (choice, yes/no noul, ordered score) in one forward pass. Apache-2.0, with MLX builds and an offline System One server.
v3 is a rank-16 LoRA, merged, trained under SemIf's direct-options prompt (credited, MIT); v2.1 stays at its own tag. Probabilities are a softmax over the option-letter logits. The maker's numbers are labelled "not an official JevBench score". It is overconfident on unfamiliar tasks (ECE 0.129 on 27 held-out Natural Instructions tasks), and the serving temperature map makes served yes/no probabilities sharp. MLX 8-bit and 4-bit builds; the serving repo speaks /v1/systemone offline.
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/Qwen3.5-4B |
| Maker | Oraculum |
| Released | 2026-09-30 |
| License | apache-2.0 |
| Reported accuracy | 85.7% |
Get the weights
pip install systemonemodels
systemone pull oraculum/manchego
The files are served from the maker's Hugging Face repository, oraculumai/Manchego, and verified against the checksums recorded here.
Read more
This page was opened by System One for Oraculum, who can claim the organisation and take it over at any time.