On August 12, 2025, Skywork AI released Matrix-Game 2.0, an open-source "interactive world model” that generates video in real time while responding to a player’s keyboard and mouse. The code and model weights are public under the MIT license. The aim: make long-horizon, controllable AI worlds something you can run, not just watch.

Unlike earlier generative video demos that produced short, non-interactive clips, Matrix-Game 2.0 is built for streaming play: it targets ~25 frames per second and can sustain minute-level sequences, binding user inputs directly into each predicted frame. The release lands a week after DeepMind’s Genie 3 preview and positions Skywork’s model as a fully open alternative for researchers and builders.

  #!/bin/bash
  # Clone repo
  git clone <REPO_URL>
  cd matrix-game-2.0
  
  # Create env & install
  python -m venv .venv && source .venv/bin/activate
  pip install -r requirements.txt
  
  # Install performance deps (GPU)
  pip install flash-attn --no-build-isolation
  pip install ninja git+https://github.com/NVIDIA/apex.git
  
  # Run: batch video
  python inference.py \
    --config configs/base.yaml \
    --ckpt checkpoints/model.safetensors \
    --vae checkpoints/vae.safetensors \
    --input prompts.txt \
    --out out/
  

python inference_streaming.py \
    --config configs/stream.yaml \
    --ckpt checkpoints/model.safetensors \
    --vae checkpoints/vae.safetensors \
    --resolution 720 \
    --target_fps 25
  

# Genie 3 — public quick start is limited
  # Typical path: request access; follow provider docs / SDK when available.

How it works (in brief)

• Few-step diffusion for real time. A distilled, auto-regressive diffusion pipeline reduces steps so the model can stream at ~25 FPS.
• Action injection. Keyboard/mouse signals are embedded at frame level, letting inputs steer each next frame.
• Data at scale. A production pipeline spanning Unreal Engine and GTA5 generated ~1,200 hours of interactive video for training.

By the numbers

• Throughput: ~25 FPS streaming
• Horizon: minute-level sequences
• Parameters: ~1.8B
• Data: ~1,200 hours interactive video
• License: MIT (open source)

Why it matters

• From demo to loop. Real-time, action-conditioned video lets teams prototype agents, UX, and simulation without hand-coding full engines.
• Reproducibility. Public weights + code mean labs can replicate, benchmark, and extend the model, rare in this corner of AI.
• Ecosystem pressure. Arriving alongside Genie 3, the release nudges the field toward open baselines for interactive worlds.

How it differs from Genie 3

• Access model: Genie 3 is in restricted preview; Matrix-Game 2.0 is open-source with runnable scripts.
• Control surface: Both are interactive; Matrix-Game 2.0 emphasizes explicit frame-level action injection.
• Public detail: Matrix-Game 2.0 documents data pipeline, dependencies, and quick start; Genie 3 shares high-level research framing.
• Performance claims: Skywork highlights 25 FPS and minute-long sequences; Genie 3 coverage highlights real-time interactivity (often cited around 720p/24 FPS) in media.

The catch

• Compute + setup. Inference depends on FlashAttention and NVIDIA Apex; practical throughput and resolution will vary by GPU. Independent, third-party benchmarks at scale are still limited.
• Generalization. Multi-agent complexity, richer physics, and long-term memory remain open research problems across world models. (DeepMind flags similar limits for Genie 3.)

Getting started (quick)

1. Clone the repo and install requirements.
2. Install FlashAttention and Apex.
3. Run inference.py for batch video or inference_streaming.py for real-time output (provide config, checkpoint, VAE path).

Outlook (near-term)

Watch for independent evaluations on the GameWorld benchmark, community ports that improve UX, and experiments coupling Matrix-Game 2.0 with agent training loops. If the open baseline holds up, expect rapid iteration—fine-tunes for domains, better latency-resolution trade-offs, and toolchains that make "AI worlds” a routine lab artifact