CommunityJev Demos & Experiments
Jevosaurus
A browser dino run steered by small decision models: each obstacle becomes a one-line text question answered through open-jev, with a three-course index comparing models in real-time or turn-based mode.
- Format
- Playground
- Published by
- Community
- Author
- shreyask
- Last verified
- Oct 2026
- Platforms
- Browser
- Capabilities
- Open sourceRuns locally
Highlights
- Each obstacle is described in one line of text and answered as a typed question through 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 rather than seconds to impact.
- The card says every small model tried answered from obstacle type alone given time to impact in seconds, jumping too early or never.
- The Jevosaurus Index replays every model move by move on the same three seeded courses, up to 60 seconds each, in real-time or turn-based modes.
- Phones get GLiNER2.5-Decide at 4-bit in 0.52 GB with a WebAssembly fallback and two questions per obstacle; ?mobile=1 tries it on desktop.
Watch out
Runs entirely in the browser: the fastest path needs WebGPU, with a WebAssembly fallback on phones at 4-bit quantization. The 70 ms figure is the card's own measurement for wording-based prompts, and the Index replays recorded runs shipped with the page rather than hosting a live competition.
Questions about this entry
Short answers drawn from the linked project or page.
When an obstacle comes near, the game is described in one line of text and a small decision model answers a typed question through open-jev: jump, duck, or keep running.
Every model plays the same three seeded courses, up to 60 seconds each, and each run is recorded move by move so the page can replay all models on the same course side by side.
Real time keeps the game running for as long as the model takes to answer, simulated per model so they never compete for the GPU; turn-based gives every answer instantly, so only judgment counts.
Because the card found decision models match text rather than compare numbers: with words like 'very close' GLiNER2.5-Decide plays indefinitely, while time-to-impact in seconds made every model tried jump too early or never.
Yes. Phones get GLiNER2.5-Decide at 4-bit in 0.52 GB with a WebAssembly path when there is no WebGPU, and per-obstacle timing pauses the game for two questions.
A MacBook Pro M3 Pro using WebGPU; that run ships with the page, while the Run it on your GPU option measures your own machine instead.
Yes, add ?mobile=1 to the URL; it enables the phone behavior of per-obstacle questions so the game plays the same at any speed.
More Jev resources like this
AutoTrust's live game-arena demo: watch JEV-27B play Mario, a 3D Rubik's Cube, and Tetris while a side panel reports action probabilities and per-decision inference time, alongside a decision lab and a ViZDoom arena.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
An interactive ZeroGPU demo of AutoTrust's JEV-9B: System 1 answers noul, choice, and score questions in one prefill pass with calibrated probabilities, while an Escalate tab hands low-confidence cases to the untouched Qwen3.5-9B base.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
A three-lane demo of pngwn's System One 4B scorer: one prefill plus a batched forward for every option branch, a naive per-option equivalence check, and a Qwen3.5-4B instruct baseline generating JSON.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
A static page that races three LLM judging gates — Laya, JEV, and zero-token ZTC — on the same Tetris board with identical pieces, proposals, and noise, publishing paired-match results and the measured constants behind them.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
A Gradio playground for Jev-Omni: give a situation plus a question with options and get one probability per option from a single forward pass, with no generated text, from a Gemma 4 12B IT fine-tune.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
A static React front-end for ra_jev_v3, described as Qwen3.5-4B plus LoRA, whose inference runs on the same organisation's ra_jev_api ZeroGPU backend Space, with mock and custom-backend URL options.
Jev Demos & ExperimentsPlayground#community#playground#demo
Community
From the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.