---
title: Jevosaurus
emoji: 🦖
colorFrom: gray
colorTo: green
sdk: static
app_file: index.html
pinned: false
short_description: A decision model plays the Chrome dino game on WebGPU
models:
  - onnx-community/GLiNER2.5-Decide-ONNX
  - onnx-community/kev-0.6b-ONNX
  - onnx-community/Decision-2.0-Kai-0.6B-ONNX
  - onnx-community/strands-decider-2B-hobson-v19-ONNX
---

# Jevosaurus 🦖

Jevosaurus is a dino run by a small decision model. It plays the Chrome dino game in your browser. Each time an obstacle comes near, the game is described in a line of text and the model answers one typed question with [open-jev](https://github.com/nico-martin/open-jev): jump, duck or keep running.

GLiNER2.5-Decide plays indefinitely at about 70 ms a decision when the prompt is written in words ("It is very close") and the options are described with the same words. Given the time to impact in seconds instead, every small model tried answers from the obstacle type alone and jumps too early (or never): decision models match text, they do not compare numbers.

## Jevosaurus Index

Every model plays the same three seeded courses, up to 60 seconds each. In **real time** the game keeps running for as long as the model takes to answer (simulated per model, so models never compete for the GPU); **turn-based** gives every answer instantly, so only judgment counts. Every run is recorded move by move, so the page replays all models on the same course side by side. The reference run (MacBook Pro M3 Pro, WebGPU) ships with the page; "Run it on your GPU" measures your own machine.

## On a phone

Phones get GLiNER2.5-Decide at 4-bit (0.52 GB; WebAssembly when there is no WebGPU) and per-obstacle timing: the game pauses for two questions per obstacle ("close": keep running, "very close": jump or duck), so it plays the same at any speed. Add `?mobile=1` to try that mode on a desktop.