Nano JEV evaluates dynamic choices.
A request supplies its own candidates. The model assigns a probability to each Choice option rather than generating a sentence as its answer.
NANO JEV / OPEN-SOURCE REPLICA
Nano JEV turns questions into probabilities.
NanoJev is TianyuCodings’ independent 0.6B open-source replica of Jev. It scores typed questions and returns probability distributions. HE Jev presents the author’s recordings and tests with attribution; this is not an HE Jev model release.
Independent project by TianyuCodings ↗ · Author-recorded tests · Not an HE Jev release
Watch the recorded testsA small model, a carefully bounded question.
A human purpose. A thoughtful next step.02 / RECORDED CASES
01 / NANO JEV / OPEN-SOURCE REPLICA
Nano JEV scores a question before code acts.
The public NanoJev project starts with a state, a typed question, and candidate answers. Its Qwen3-0.6B backbone and decision heads produce distributions; a separate controller can use those judgments to choose the next move.
Nano JEV returns distributions.
The project also records Boolean and ordered Score judgments. A distribution gives the surrounding application more to inspect than a single hard label.
Code carries the longer task.
In the recorded games, shared code tracks state, checks possible moves, and executes actions. NanoJev supplies focused judgments; the controller handles planning and game rules.
02 / RECORDED CASES
Nano JEV shows the judgment and the journey.
These are selected recordings from the project’s public test artifacts. The author retained the actions and outcomes; the clips are not a live model on this page or an average across all runs.
ARCHIVE / 01
01 / FIXED 50 × 50 MAZE
Nano JEV reaches the exit.
244 attempts · 36 collisions · goal reached
The model judges four local directions. Shared code remembers collisions and explores physically verified paths. The 50 × 50 case is a fixed out-of-distribution showcase, not a broad navigation score.
Original video ↗ARCHIVE / 02
02 / SELECTED 12 × 12 SNAKE
Nano JEV survives the horizon.
27 food collected · 256 steps · alive at horizon
The shared planner filters immediate collisions and finds food paths. The model breaks ties among remaining actions; 181 of 256 moves in this recording were forced by code. This is an outcome-selected highlight.
Original video ↗03 / Explore the JEV family
Explore the JEV approaches.
Choose the page that matches your question: applications, model thinking, language, integration, or smaller tasks.