vera-core
A 1.9B-parameter decision model from the scar-ai account: a Qwen3.5-2B backbone with its own decision head reads a state and typed questions (choice, score, yes/no) and returns a probability per option in one forward pass, generating no text. 8K context, Apache-2.0.
The Qwen3.5-2B backbone gets a purpose-built head that scores each question's options, and the whole stack is fully fine-tuned on decision data the card does not describe. Choice takes 2 to 255 options, score 2 to 10 levels. It speaks Jev's /v1/systemone shape through the maker's vera-serve (no auth, binds 127.0.0.1); it needs custom code or the vera-s1 package, and Apple MPS is not supported. The maker presents it as a generalist base to fine-tune. On JevBench's 231 public items (a third-party suite; the maintainers have not ranked it) the maker's own run gives 77.9% (easy 100, original 93.1, hard 58.6) with ECE 0.098; that is the only calibration evidence. Sibling vera-spark (about 638M, ModernBERT-large with MoE layers, same interface) scores 68.8% in the same run.
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 | 1.9B |
| Base model | Qwen/Qwen3.5-2B |
| Maker | scar-ai |
| Released | 2026-10-03 |
| License | apache-2.0 |
| Reported accuracy | 77.9% |
| Reported latency | 41 to 44 ms median per question on an MI300X, batch of one |
Get the weights
pip install systemonemodels
systemone pull scar-ai/vera
The files are served from the maker's Hugging Face repository, scar-ai/vera-core, and verified against the checksums recorded here.
Read more
This page was opened by System One for scar-ai, who can claim the organisation and take it over at any time.