Does the model actually see the game screen?

No. It reads only symbolic text state plus one instruction sentence per scenario and the action descriptions, which is the point: decisions from structured state, not vision.

Why is Doom relevant to my ticket router or moderator?

Doom is the fastest way to watch a decision model think; the same choice, score, and noul questions route tickets, moderate comments, and pick agent UI actions.

What if the hosted API route is not live yet?

Use the local-weights backend or point the URL backend at your own server speaking the same POST /v1/systemone contract; the harness also works against any System One model.

Will remote play react too slowly over the internet?

Yes, remote decisions run longer per step than local GPU inference, so the script supports raising the tick count to give the marine more time per decision.

What does the model actually see each step, and what can it do?

Text only, and it answers two typed questions in one forward pass. The state is health, ammo, kills, crosshair status, enemies with distance and bearing plus visibility, items, wall depth ahead and recent actions, plus one instruction sentence per scenario and the 7 action names with descriptions. The answers are a choice among 7 actions with a probability per action, and a noul for whether an enemy is in the crosshair.

How do I install and run this, and what are the scenario options?

Clone the repo, make a virtual environment, activate it and pip install -r requirements.txt plus opencv-python for the live HUD window. There are seven ViZDoom scenarios: deadly_corridor by default, defend_the_center, defend_the_line, health_gathering, basic, my_way_home and take_cover. Flags cover skill 1 through 5, a seed, --lang en or th, --seconds to stop automatically, --record to save a clip with ffmpeg, and --no-window for headless use.

What license is this under, and what was it modelled on?

The model is iapp/OpenThai-SystemOne by iApp Technology and OpenThaiGPT, Apache-2.0, and the game is ViZDoom. It was inspired by TypeSafe's Jev Doom demo and NanoJev. Because the request and response format matches TypeSafe's /v1/systemone, the same harness also runs against any other System One model rather than only this checkpoint.

CommunityJev Models & Reimplementations

Openthai Systemone Doom

Tiny open 0.8B decision model with no text generation or vision that plays Doom in real time from symbolic game state, answering movement and crosshair questions in one forward pass.

See highlights

Open github.com

Stars: 1
Language: Python
License: APACHE-2.0
Published by: Community
Author: iapp-technology
Last verified: Oct 2026

iapp-technology/openthai-systemone-doom

Organization repository on GitHub

View on GitHub

OpenThai-SystemOne plays Doom: a 0.8B decision model controlling ViZDoom in real time (iApp API or local weights)

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A tiny open decision model (0.8B, no text generation, no vision) controlling Doom in real time, on your own machine.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Sends health, ammo, enemies, depth, and recent actions as text; the model picks one of 7 actions and reports crosshair state per step.
Runs near 40 ms per decision on a GPU with zero output tokens; remote API play takes longer and supports a slower tick option.
Three ways to run: hosted API key, local open weights on GPU or Apple Silicon, or any server speaking the same systemone contract.
Ships ViZDoom scenarios with skill, seed, HUD language, recording, and headless flags; the model never saw Doom in training.

Watch out

Game demo over an open model; gameplay skill is illustrative, not a capability claim. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

gut

Python helper that turns judgment calls into one-liners like likely, classify and rate, defaulting to Jev while also running on local or OpenAI-compatible backends.