NeoHorse-Jev-4B
TokenRhythm's prefill-only decision model for agent workflows. Built on its NeoHorse-1-4B (a Qwen3.5-4B derivative), it predicts Choice, Noul and Score answers over application-defined options without generating text, with optional single-image input.
Runs on vLLM, SGLang or a bundled Python, CLI and HTTP runtime (wheel included). TokenRhythm reports 75.3% per-example accuracy on JevBench's 231 public items, 81.9 on the Kev suites, 87.2% on a 282-item Nimble subset and a 77.70 equal-weight mean over six text suites, plus 60.7% on Image-NLI. The card says it has not published latency, memory or cost comparisons under a common protocol. A GGUF build and a ModelScope mirror exist. The third-party Decision Index 0.2.1 scores it 36.75.
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 calibrated probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 4B |
| Base model | tokenrhythm/neohorse-1-4b |
| Maker | TokenRhythm |
| Released | 2026-09-23 |
| License | apache-2.0 |
| Reported accuracy | 75.3% |
Get the weights
pip install systemonemodels
systemone pull tokenrhythm/neohorse-jev
The files are served from the maker's Hugging Face repository, TokenRhythm/NeoHorse-Jev-4B, and verified against the checksums recorded here.