Certo Unified 4B
AltSlate Labs' research-preview decision model on Qwen3-4B. It reads a state and a list of described options and returns a calibrated probability per option for choice, score and yes/no (noul) questions, read by a small head at the last prompt token.
A research preview. On its fast path a small head reads the last prompt token; an optional reasoning path writes a short rationale first, which is not one pass. On AltSlate's own held-out evaluations choice reaches 0.93 to 0.97 and noul 0.88 to 0.92, with temperatures fitted on a separate split; score on an unseen rubric is about 0.46. On the third-party JevBench public items the maker's own run gives 0.71 with ECE 0.22. head.pt is a PyTorch pickle. Family members: certo-decision-model v1 (MIT, an encoder trained on Certo's synthetic world), certo-decision-v2.1 (MIT) and certo-r1-qwen3-4b (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 |
| Maker | AltSlate Labs |
| Released | 2026-09-28 |
| License | apache-2.0 |
Get the weights
pip install systemonemodels
systemone pull altslate-labs/certo
The files are served from the maker's Hugging Face repository, altslate/certo-unified-4b, and verified against the checksums recorded here.
Read more
This page was opened by System One for AltSlate Labs, who can claim the organisation and take it over at any time.