NanoJev

# NanoJev — A nano replica of [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)

**English** | [简体中文](README.zh-CN.md)

**A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.**

> **New project: [JevHarness](https://github.com/TianyuCodings/JevHarness)** — Let an LLM build task-specific decision harnesses with [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), with optional refinement using rewards and execution traces. Includes an interactive Pokémon demo.

[Play ViZDoom](https://nanojev-dev.tianyuchen99.chatgpt.site/?autoplay=1) · [Maze & Snake](https://nanojev-dev.tianyuchen99.chatgpt.site/side-by-side?autoplay=1#maze) · [Model](https://huggingface.co/C-Tianyu/NanoJev) · [Dataset](https://huggingface.co/datasets/C-Tianyu/NanoJev-Data)

**Now playing ViZDoom:** one shared checkpoint handles Basic aiming and Predict Position's moving-target rocket shots, alongside Maze and Snake.

**4 tasks · 18,760 decision questions per data variant · 896 Predict Position expert episodes**

## What's new

**September 20, 2026 — One model, four games.**

- **ViZDoom Basic:** **128/128** test successes, compared with **56/128** for [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev).
- **ViZDoom Predict Position:** **27/128** test successes, up from **11/128** before this round; the matched [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) run also scores **11/128**. The policy learns when to turn, wait and fire at a moving target.
- **16,333 ViZDoom questions** within an **18,760-question** mixed-task dataset per target variant, spanning train, dev, calibration, test and OOD.
- **One model, four games:** the same step-400 checkpoint also