Overview & result
Recreates Jev's parallel-decision approach so a 0.6B small model outputs the full probability distribution in one forward pass, without decoding output tokens. github.com/TianyuCodings/…
Show the original post (Chinese)
复刻 Jev 的并行决策思路，让 0.6B 小模型一次前向直接输出完整概率分布，不做输出 token 解码。
github.com/TianyuCodings/…
How Jev fits in the loop
- NanoJev is a 0.6B model that takes a state and a set of questions.
- It returns a complete probability distribution for each question in parallel, with no output-token decoding.
- One shared checkpoint plays four tasks: ViZDoom Basic aiming, Predict Position, Maze and Snake.
- Browser replays compare Jev, NanoJev and an untuned Qwen model on the same tasks.
How to reproduce
- Inspect the original showcase and source material at https://github.com/TianyuCodings/NanoJev
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
It is a small open replica that lets people study the parallel-decision approach on game tasks.
Limitations
- A community replica, not TypeSafe's model.
- The README shows it on four game tasks only.
- The interactive development site currently requires access; recordings can be played locally.