solvi-large (preview)
A 396M ModernBERT-large cross-encoder that answers typed questions about a text or a JSON state: one or several options, scores, yes/no with "not stated", spans with evidence quotes, rankings and numeric bins, each with a confidence and an act-or-escalate signal.
Published by Maxim Kuznetsov as a preview, backing his Apache-2.0 solvi runtime. One question per forward pass; option lists are given when the question is asked; spans are pointed to in the text, not generated. On Fastino's public Fast Decisions dev split, run in the maker's own harness, it scores 59.4% zero-shot and 63.0% after 64 labelled examples; GLiNER2.5-Decide scored 62.9% in the same harness, and the maker says solvi-large does not beat it on zero-shot choice. Zero-shot confidences are not calibrated (ECE 0.21), and the shipped act threshold is over-confident on new text. English only; 512 tokens. Siblings solvi-base (150M, distilled) and solvi-large-long (8,192 tokens). The training-data manifest excludes non-commercial, research-only and unlicensed data; three sets are share-alike.
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
- rank — orders a set of items
- extract — pulls spans or fields out of the input
At a glance
| Parameters | 396M |
| Base model | answerdotai/ModernBERT-large |
| Maker | solvi |
| Released | 2026-09-28 |
| License | apache-2.0 |
| Reported accuracy | 59.4% |
| Reported latency | 137 ms per short question on a CPU (ONNX fp16) |
Get the weights
pip install systemonemodels
systemone pull solvi-ai/solvi
The files are served from the maker's Hugging Face repository, solvi-ai/solvi-large, and verified against the checksums recorded here.