MIMIR-1
A non-generative decision model: a ModernBERT-Large encoder (419M) with a decision head. Context, question and options go in; a probability per option comes out, with ABSTAINED when no option fits and DEFERRED below certified risk thresholds. English only.
Published under the Mythologic name by Abderahmane Ainouche, with an Apache-2.0 SDK (mimir-decisions on PyPI). Its decision types are choose, multi-choice, yes/no, verify (supported, contradicted or insufficient), rank, rate and bounded estimate, and each answer names the context passages behind it. The card reports selective-risk thresholds certified on 95,550 held-out records rather than a calibration error. Multilabel and continuous answers defer at every risk level in this release, and inputs are cut into 512-token chunks. In the maker's own runs it scores 0.883 on Banking77, 0.866 on MASSIVE en and 0.840 on BoolQ, and trails on fast-decisions (0.432) and DAIR Emotion (0.080, abstaining on 1,792 of 2,000 records). The MIMIR Model License allows use, including commercial use, below US$1M group revenue; hosting it as a service or redistributing it as a product needs a commercial agreement, so this registry links to the weights and does not mirror or serve them.
What it decides
- choice — picks one option from a set
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- rank — orders a set of items
- score — places the input on an ordered scale
- route — sends the input to one of several destinations
At a glance
| Parameters | 419M |
| Base model | answerdotai/ModernBERT-Large-Instruct |
| Maker | Mythologic |
| Released | 2026-09-28 |
| License | MIMIR Model License 1.0 (free below US$1M revenue; hosting needs a deal; non-commercial for us) |
| Reported accuracy | 72.5% |
Get the weights
pip install systemonemodels
systemone pull mythologic/mimir
The files are served from the maker's Hugging Face repository, Mythologic/MIMIR-1, and verified against the checksums recorded here.