NanoJev
0.6B open Jev replica returning full option distributions with no decoding; one checkpoint plays ViZDoom, Maze and Snake.
Overview
NanoJev is a 0.6B parallel decision model inspired by Jev: states and questions go in, complete probability distributions over candidates come out, with zero output-token decoding. It ships an end-to-end training pipeline, weights on Hugging Face (C-Tianyu/NanoJev) and a dataset, and supports dynamic candidate sets.
The September 2026 release uses one step-400 checkpoint across four games. The README reports 128/128 test successes on ViZDoom Basic against 56/128 for Jev on the author's split, 27/128 on ViZDoom Predict Position against 11/128 for Jev, a 50x50 maze completed in 225 attempts, and 30 food items in a 256-step Snake run. The dataset has 18,760 decision questions per variant across train, dev, calibration, test and OOD. Browser replays compare Jev, NanoJev and untuned Qwen. The author's follow-up JevHarness project builds task-specific decision harnesses. It suits researchers and game-AI hobbyists wanting a small, trainable decision model.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action should the agent take next in this game state? | — |
How it works
- Encode the game state and candidate actions as a decision question.
- Run a single forward pass to get a distribution over candidates.
- Take the argmax action (code can pre-filter candidates).
- Train or fine-tune with the included pipeline.
Strengths
- Small 0.6B model with open weights, data and training code.
- No decoding, so outputs are always valid candidates.
- Recorded side-by-side comparisons against Jev.
Caveats
- Game results are on the author's own splits.
- Interactive development site requires access; local replays work.
- Trained mainly on game tasks.
Getting started
Download the model from https://huggingface.co/C-Tianyu/NanoJev and follow the README to run the decision server and local replays.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch, local decision server
- Has code
- yes
- License
- MIT
- Stars
- 2,543
Use cases and integrations
How curated lists describe it
Open replica: a 0.6B parallel decision model that returns full probability distributions with no output-token decoding, shipped with its training pipeline, weights, and dataset.
yibie/awesome-jev0.6B parallel decision model: state and questions in, a full distribution out, no decoded text. Published weights and one checkpoint for ViZDoom, Maze, and Snake; on the author's ViZDoom Basic split, 128/128 against 56/128 for Jev.
AnotiaWang/awesome-decision-modelsNano replica of Jev featuring parallel decisions, dynamic candidate sets, and an end-to-end training and evaluation pipeline.
logicrw/awesome-jev-projectsA nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.
wh000wh000/awesome-jev-liveSmall parallel-decision model with dynamic candidates, a training pipeline, and recorded game comparisons that include shared code planning.
cobanov/awesome-jevA nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.
hellogumbo/awesome-jev0.6B parallel decision model with an end-to-end training pipeline.
MrJev/awesome-jev
Frequently asked questions
How big is NanoJev?
0.6B parameters.
How does it compare to Jev on games?
The README reports 128/128 versus 56/128 for Jev on its ViZDoom Basic split.
Is the training pipeline included?
Yes, along with weights and the dataset.
Related projects
Browser arcade games (2048, runner, chess, Flappy Bird, Sudoku) where small decision models choose each move from text options.
ProjectprototypePython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Python library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePython33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonPre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPythonReads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.
ProjectresearchPython