Misc

Getting Started with VAST's Open-Source EdenLive World Model

A guide to EdenLive, the real-time world model VAST open-sourced in Oct 2026: the 5B model exceeds 24FPS at 720P on one RTX 5090; steps, inputs and officially listed limitations

One-sentence summary

VAST released and open-sourced EdenLive, a real-time world model, in October 2026: according to the company, its 5B-parameter model generates at over 24FPS at 720P on a single NVIDIA RTX 5090, supports reference scenes or videos, a target first frame and prompts; model weights, inference code, a technical report and demos are open.

Use Cases ​

  • Real-time world-model applications that start from a reference scene or video and continuously generate minute-scale dynamic frames.
  • Developers running world-model inference on a single consumer GPU (RTX 5090 class).
  • Researchers evaluating video-generation models across input modes (reference frames, first frames, prompts).

Steps ​

  1. Get resources: obtain the EdenLive model weights, inference code, technical report and demos from VAST's official open-source release (public since October 2026).
  2. Check hardware: per the company, the 5B model runs on a single NVIDIA RTX 5090, generating at over 24FPS at 720P.
  3. Prepare inputs: use a reference scene or video as the starting point, optionally specify a target first frame, and provide prompts to steer generation.
  4. Run inference: load the weights with the official inference code and generate, verifying frame rate and resolution.
  5. Use the interaction system: the upstream interaction system maintains world state and rules for longer, consistent generation.
  6. Record limitations: check the officially listed limits for high-frequency interaction, multi-character identity consistency, fine control of small objects and long-horizon visual consistency.

Notes ​

  • Performance figures (24FPS, 720P, single RTX 5090) follow VAST's official disclosure; actual results vary with hardware and inputs.
  • Officially listed limitations include: high-frequency interaction, multi-character identity consistency, fine control of small objects, and long-duration visual consistency.
  • Weights and code come from the open-source release; deploy per the official technical report, and cite source and publication date when referencing data.

References ​

  • e/acc AI Daily (October 19, 2026; original report: QbitAI): VAST open-sources EdenLive world model with over 24FPS generation at 720Plink