TianyuCodings/NanoJev

A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.

What it solves

NanoJev is a small-scale replica of the Jev model designed to handle decision-making tasks without the need for output-token decoding. It allows a model to take in states and questions and output complete probability distributions over possible actions or outcomes, enabling faster and more direct decision-making in gaming environments like ViZDoom, Maze, and Snake.

How it works

Built on a Qwen3-0.6B backbone, NanoJev uses specialized decision heads to process requests containing a state, a question, and a set of candidate paths. Instead of generating text, it uses these heads to produce direct probabilities:

  • Choice head: Uses set attention and softmax to return a distribution over 2–255 candidates.
  • Boolean head: Uses a sigmoid to predict the probability of a proposition.
  • Score head: Returns a probability-weighted level (2–10 ordered levels).

Who it’s for

It is intended for researchers and developers interested in parallel decision models, efficient AI agents for games, and those exploring alternatives to traditional token-based LLM decoding for action selection.

Highlights

  • Zero-token decoding: Ranks and selects actions via direct probabilities rather than generating text answers.
  • Multi-task capability: A single 0.6B parameter checkpoint handles four different games (ViZDoom Basic, ViZDoom Predict Position, Maze, and Snake).
  • High efficiency: Outperforms larger models in specific tasks, such as eliminating targets in ViZDoom Basic in 1.40 seconds.
  • Comprehensive dataset: Includes a mixed-task dataset of 18,760 decision questions and expert gameplay episodes.

Related

  • Project
    jaredpalmer/kev

    A family of small, self-hostable decision models based on Qwen that provide calibrated probabilities for yes/no, multiple-choice, and rating tasks.

  • Project
    malevrigns/agent-jev

    AgentJev is a 0.6 B specialist model that, given a JSON‑encoded state and typed questions, returns probability distributions over supplied options without decoding any tokens. Built on Qwen‑3‑0.6B with a custom candidate head, it provides fast, token‑free decisions (boolean, choice, score) for agent loops, outperforming comparable models on the Typed‑Decisions benchmark and offering shared‑prefix caching for wide‑candidate loads.

  • Project
    Anil-matcha/awesome-jev-by-typesafe

    A collection of use cases and patterns for Jev, a System One model by TypeSafe AI that provides typed decisions (Choice, Score, Noul) instead of generative text for software control flow.

  • Project
    razorback16/openjev

    An open-source decision server that uses DiffusionGemma to provide fast, calibrated, and strictly typed answers (yes/no, choice, score) by reading model probabilities instead of generating text.

  • Dispatch
    Comparison of Non-Autoregressive Probability Prediction Models: Jev and Early Open-Source Implementations

    An analysis of the architectural similarities between the Jev model and earlier open-source work by nandakishor_ml, focusing on non-autoregressive probability prediction and structured output.