Skip to main content
Back

Tencent Opens Hy-Nomas to Build Steerable Video Worlds on One GPU

Tencent Hunyuan has released Hy-Nomas, an open interactive video world model that generates 720p footage one segment at a time under live text and camera control.

  • Tencent released Hy-Nomas, an Apache 2.0 interactive video world model, on Hugging Face.
  • Generates 720x1280 video at 24 FPS using only 8 distilled DiT denoising steps per segment.
  • Segment-autoregressive MMDiT with 54 blocks, Qwen2.5-VL text encoder, and camera cross-attention.
  • Single-GPU RTX 4090 mode via INT8 DiT, INT8 KV cache, FP8 PixelUMM decoder, and CPU offload.
  • Ships with Hy-Nomas-Bench, 280 test cases for multi-turn text, camera, and spatial memory.
  • Includes cinematic and dual-arm robotics LoRAs plus a free hosted interactive demo.

Tencent’s Hunyuan team has released the Hy-Nomas weights and source code under Apache 2.0. The model generates a continuing video world in five-second segments, accepting new text instructions and camera trajectories before each segment. Persistent attention state carries prior context forward, allowing earlier objects, layouts, and actions to influence later video.

At publication, Hugging Face recorded roughly 88,000 downloads. The primary configuration generates 720 × 1,280 video at 24 frames per second using eight denoising steps per segment. Tencent also provides a quantized configuration designed to fit on one RTX 4090, giving developers a local route to test multi-turn world generation without an eight-GPU server.

Generation keeps moving

In this release, “world generation” means extending and redirecting generated video over multiple turns. Each instruction can introduce objects, change an environment or visual style, describe an action, or continue an existing event. Camera controls can alter the viewpoint and motion during the same sequence.

  • Text interaction: Successive prompts control subjects, actions, appearance, style, and environment.
  • Camera control: Trajectories specify viewpoint and movement alongside the text prompt.
  • Persistent context: A key-value cache retains attention features from completed segments.
  • Segment streaming: The hosted demo returns decoded output as segment boundaries are reached.

The cache contains the keys and values computed by the model’s attention layers. Reusing those tensors gives later segments access to prior visual context without denoising the full video history again. Continuity remains probabilistic because the cache stores learned features rather than explicit geometry, object records, or physics state.