Enigma

[Enigma @Enigma_AI]

So we did a thing

[Jonathan Jacobi @j0nathanj]

Introducing Multiverse: the first AI-generated multiplayer game.

Multiplayer was the missing piece in AI-generated worlds — now it’s here. Players can interact and shape a shared AI-simulated world, in real-time.

Training and research cost < $1.5K. Run it on your own PC.

We @enigma_ai are open-sourcing everything: code, data, weights, architecture, and research.

Here’s a closer look (and how it works!):

[Jack Parker-Holder @jparkerholder]

World models could lead to a significant leap for open-ended learning and embodied AGI. Multiplayer brings another dimension. Great work!!

[Chubby♨️ @kimmonismus]

Multiplayer in an AI simulated world. This has become reality faster than expected. It's absolutely fantastic what is possible here.

[Rohan Paul @rohanpaul_ai]

Cool work by @enigma_ai .

Open-sourced Multiverse, the first AI multiplayer world model, built for under $1.5K.

It simulates real-time shared environments using a novel architecture combining diffusion and joint-action modeling. Everything��code, data, weights, training pipeline—is publicly released.

→ Multiverse extends traditional single-player world models by combining both players’ inputs into a single joint action embedding and processes both frames as one unified entity using channel-axis stacking for consistency.

→ Uses a U-Net-based diffusion model for simultaneous dual-frame generation, ensuring both players see coherent and causally consistent outcomes in real-time.

→ Captures vehicle kinematics and low-speed relative motion using a clever sparse frame sampling strategy—latest 4 frames + every 4th frame over the next 4—spanning 0.666 seconds of context.

→ Trained on modified Gran Turismo 4, with custom 1-v-1 mode using replay footage and computer vision to extract control inputs from HUD indicators.

→ B-Spec mode enabled fully autonomous data generation, massively scaling the dataset without human input logging.

→ Achieved 15-second autoregressive prediction with curriculum learning and memory-efficient paged frame loading to manage GPU VRAM constraints.

[GTPlanet @GTPlanetNews]

Gran Turismo 4 has been used to train a new AI research project called "Multiverse" by @enigma_ai. The team of AI researchers (and Gran Turismo fans) have developed a new technique for creating AI-generated worlds for multiplayer games: https://t.co/PlMNN0TsR7

[Enigma @Enigma_AI]

echo "Enigma is thrilled to announce: the first AI-generated multiplayer game. Gran Turismo 4. Play it on your own PC" | sha256sum

[Gal Niv @GalNivs]

ba059d5ad7b8dc45c175a382f479368833666ef189056404f719c5627e68ffc3

[clem 🤗 @ClementDelangue]

This is the coolest dataset I've seen on @huggingface today: action labels of 1v1 races in Gran Turismo 4 to train a multiplayer world model!

Great stuff by @enigma_ai!

[Enigma @Enigma_AI]

Us 🤝 Datasets

[clem 🤗 @ClementDelangue]

This is the coolest dataset I've seen on @huggingface today: action labels of 1v1 races in Gran Turismo 4 to train a multiplayer world model!

Great stuff by @enigma_ai!

[Enigma @Enigma_AI]

Multiplayer: ✅

[Colin Campbell @ccamp___]

Across 3D and video, there are many research challenges in World Models being tackled by startups and top research labs alike:

- Maintaining state
- Long term memory
- Precise controllability
- Accurate physics
- Multimodal input
- Multiplayer
- Reasoning
- ...

[AK @_akhaliq]

Enigma labs dropped Multiverse on Hugging Face

AI Multiplayer World Model

Multiplayer was the missing piece in AI-generated worlds — now it’s here. Players can interact and shape a shared AI-simulated world, in real-time. Training and research cost < $1.5K. Run it on your own PC.

[Jonathan Jacobi @j0nathanj]

Introducing Multiverse: the first AI-generated multiplayer game.

Multiplayer was the missing piece in AI-generated worlds — now it’s here. Players can interact and shape a shared AI-simulated world, in real-time.

Training and research cost < $1.5K. Run it on your own PC.

We @enigma_ai are open-sourcing everything: code, data, weights, architecture, and research.

Here’s a closer look (and how it works!):