Introduction

WanGP v18.81, released October 25 (UTC) by deepbeepmeep, adds a "One World Generator": a Worlds gallery where you build an explorable world from text with Tencent's Hy-Nomas 8B or from a single image with AlayaLab's EVOKE Turbo 14B, then walk it with your keyboard. It runs locally, saves your exploration so you can pick it up later, records the walk as video, and, per the release notes, starts at 8GB of VRAM.

We covered the Hy-Nomas launch and its 12GB ComfyUI port last week. This is the next step down in hardware, and the first time an open text-to-world model and an image-to-world model sit side by side in one app built for exploring them.

What shipped

  • Tool: WanGP v18.81 "The One World Generator" (open-source local generation app, about 11k GitHub stars)
  • Author: deepbeepmeep
  • World models inside: Hy-Nomas World Model 8B (Tencent) and EVOKE Turbo World 14B (AlayaLab)
  • Modality: text or image to a continuing, keyboard-steered video world
  • Access: open weights, run locally. Both world setups are non-commercial (see the licensing note below)
  • Where to run it today: WanGP on your own NVIDIA or AMD GPU, or the 1-click Pinokio launcher
  • Weights: prepared checkpoints in DeepBeepMeep/Worlds on Hugging Face, downloaded automatically the first time you open a world model

What the Worlds gallery does

  • Walk it with the keyboard. Move and turn the camera through the generated place instead of writing a camera path first.
  • Save and continue. Store an exploration and resume it later, so a world isn't lost when you close the app.
  • Record the journey. Export your walk as a video.
  • Capture a view. Grab a frame and send it into another WanGP generation, for example as a start image for a video or 3D model.
  • Steer it with your built-in assistant. WanGP's Deepy Prime assistant can now "generate 3D assets and worlds, or continue a saved world from movement instructions" before you take over.

Two models, two starting points

Hy-Nomas 8B starts from text. Describe a place, then walk around, turn, or type an event mid-run. The release notes use "the fox sits down" as the example. It remembers earlier views to keep places and characters consistent, and you choose how much history it keeps, trading quality for quicker moves. WanGP uses Tencent's distilled INT8 transformer unchanged.

EVOKE Turbo 14B starts from an image. It keeps a depth memory of the scene, built with ViGeo 1.1, so streets and rooms hold their layout when you come back to them. The notes pitch it at streets, landscapes, and interiors at 640×384, and warn that each move takes longer than with Hy-Nomas.

Practical split for game work: use Hy-Nomas when you're writing a level from scratch and want to try events, and EVOKE when you already have concept art or a screenshot and want to see what's around the corner.

Hardware

The 8GB floor comes from WanGP's new MMGP v4 memory manager, which the author says costs "almost" no speed. The same notes are blunt about expectations: "Don't expect miracles," and near-real-time exploration depends on the model and a fast, recent GPU. Community tester cocktailpeanut expects Hy-Nomas to run in real time on 24GB+ cards but probably not on 8GB.

For the speed-ups, the release also asks you to update WanGP's optional kernels (WGPKernels 1.4.2, GGUF kernels 1.0.29) using the installation guide.

Install or update

New install on RTX 20 to 50 series (CUDA 13.1), from the repo's installation guide:

git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP
conda create -n wan2gp python=3.11.14
conda activate wan2gp
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt
python wgp.py

Already running WanGP? git pull inside the folder, then pip install -r requirements.txt. GTX 10 series and AMD (RDNA 2 to 4) have their own sections in the guide. If you'd rather skip the terminal, Pinokio has a 1-click WanGP launcher.

Licensing note

Read this before you use a world in a shipped project. The transformer weights are Apache-2.0, but each world needs a non-commercial part to run. EVOKE's depth memory uses ViGeo 1.1, which is CC-BY-NC-4.0. Hy-Nomas worlds are decoded by NVIDIA's PixelUMM decoder, under the NVIDIA One-Way Noncommercial License (research and evaluation only). The DeepBeepMeep/Worlds card states it plainly: "Both models are therefore for non-commercial use only." Treat the output as previs and pitch material, not final game content.

Conclusion

A few days ago, steering an open world model meant Docker configs, then a 12GB ComfyUI graph. Now it's a gallery in an app many local creators already have installed, with save files, recording, and an image-to-world option alongside text. It's still slow, short on resolution, and non-commercial, so it isn't a game engine. But for blocking out a level idea, walking a piece of concept art, or showing a team what a place might feel like, an 8GB card is now enough to try.

—Titus