Tencent Opens Hy-Nomas, an Interactive Video World Running on One GPU
Tencent Hunyuan open-sources an 8B diffusion video world model you can steer with text and camera moves across multiple turns.
- Tencent Hunyuan open-sourced Hy-Nomas, an 8B interactive video world model under Apache 2.0.
- Supports free-form multi-turn text prompts and continuous camera control across autoregressive video segments.
- Outputs 720p at 24 fps, up to 60 seconds, distilled to eight denoising steps.
- Runs on a single RTX 4090 using INT8 weights, INT8 KV cache, and CPU offload.
- Ships with weights, cinematic and robotics LoRAs, Nomas-PE prompt enhancer, and Hy-Nomas-Bench (280 cases).
- Beats competing open world models in human evaluations on text, camera, and spatial memory.
Tencent’s Hunyuan team has released Hy-Nomas, an interactive video world model that accepts new text instructions and camera paths as a scene develops. It generates video in connected segments while preserving objects, visual style, and spatial layout across the session.
The Apache 2.0 release includes inference code, training code, model weights, adaptation tools, and a 280-case benchmark. Tencent also provides an INT8 configuration that fits on one RTX 4090 by combining quantized weights, a reduced-precision decoder, CPU offload, and a sliding attention window.
What Tencent shipped
| Component | Details |
|---|---|
| Backbone | 8-billion-parameter diffusion transformer |
| Output | 1280 × 720 video at 24 frames per second |
| Session length | Up to 60 seconds at 720p |
| Controls | Multi-turn text instructions and camera trajectories |
| Inference | Distilled eight-step denoising process |
| Local hardware | Single RTX 4090 for the memory-optimized INT8 configuration |
| License | Apache 2.0 |
A generated world with continuity
Hy-Nomas organizes generated video as a persistent session. A user supplies the opening description, receives the first segment, and then continues the scene with instructions such as moving the camera, changing a character’s action, or revisiting an earlier location. The model uses prior segments as context when producing each new one.
DeepMind’s Genie 3 provides a prominent reference point for this category of interactive world model. Tencent compares Hy-Nomas with Genie 3 in its research paper, while releasing the weights and implementation needed to inspect and adapt its own system.
A memory built from segments
The model’s autoregressive unit is a multi-frame segment. Each new segment begins as Gaussian noise in a compressed latent representation, which stores video in a lower-dimensional form that requires less computation than raw pixels. The diffusion transformer gradually converts that noise into the next piece of video.
During generation, a key-value cache stores intermediate attention representations from earlier segments. Those cached values help the model recover previously seen furniture, lighting, objects, and viewpoints when the camera returns to an earlier location. Full-history configurations can expose more of the session to the model, while the RTX 4090 setup limits active history with a sliding window.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
