MATILDA-jev v1
Maincode's one-pass decision model: a 26.1B checkpoint whose 255-option readout scores the options of choice, noul (yes/no) and ordered score questions over text, JSON or image state, with a stored calibration temperature and no text-generation head.
The card does not name a base model; the modeling code subclasses Transformers' Qwen3.5 classes and the LICENSE carries Alibaba Cloud's Apache-2.0 notice. The weights were replaced with "v1.3" on 2 October; the card's validation (25 of 25 smoke questions, exact probability parity with the earlier checkpoint, on an AMD MI355X) dates from 30 September and predates that swap. Maincode quotes a Decision Index 0.2.1 score of 59.59 (raw 69.18, breadth 58.39) from its own run over 150,317 requests, awaiting official submission; the Decision Index is a third-party board, and the model was not on its public board as of 6 October. No accuracy, calibration-error or latency figures are given. BF16 weights need about 49 GiB; load with trust_remote_code or the bundled maincode_jev_serve runtime (POST /v1/systemone); only AMD MI355X has been tested. A leftover line on the card still mentions the repository being private; it is public and ungated.
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 | 26.1B |
| Maker | Maincode |
| Released | 2026-09-30 |
| License | apache-2.0 |
Get the weights
pip install systemonemodels
systemone pull maincode/matilda-jev
The files are served from the maker's Hugging Face repository, Maincode/matilda-jev-v1, and verified against the checksums recorded here.
Read more
This page was opened by System One for Maincode, who can claim the organisation and take it over at any time.