What it is
NanoJev is an open replica of Jev: a 0.6B parallel decision model that takes states and questions and returns complete probability distributions without output-token decoding, shipped with its training pipeline, weights, and dataset for four game tasks.
How it uses Jev
Each request supplies a state, a question, and its candidates. The backbone encodes candidate paths; shared heads produce requested probabilities. Choice uses set attention and softmax, Boolean uses sigmoid, and Score returns a probability-weighted level. Results rank, select, or sample actions without generating answer tokens.
Primitives:choicescorenoul
Technique worth stealing
Batch independent states, questions, and candidate paths in one backbone forward for parallel decisions.
Try it
Clone repo, install requirements-toy.txt and huggingface_hub, download checkpoint and data, run scripts/serve_decisions.py.
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricyes0.82
- Shows a System One patternunclear0.60
- Handles uncertaintyno0.04
- Measuredno0.15
- Runnableunclear0.43
- Worth recommendingno0.36
- Model replicayes0.97
- Problem scopescore on a 0–2 scale0.05
- About Jevyes0.97
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.