Hy-Nomas, Tencent's new open-source world model you can steer with text

Hy-Nomas is Tencent’s open-source model for generating interactive video worlds that can be steered as they unfold. Users can issue new text instructions between video segments to add objects, change scenes or styles, and describe actions; camera trajectories can also guide viewpoint and movement. A segment-autoregressive diffusion transformer generates each segment while a persistent key/value cache carries forward context, supporting continuity across turns. Text comes from a Qwen2.5-VL encoder, camera motion is encoded as conditioning, and a fine-tuned PixelUMM decoder turns compressed video latents into pixels. The distilled setup uses eight denoising steps and produces 720 × 1280 video at 24 frames per second.

The release includes Hy-Nomas-Bench, a 280-test evaluation suite covering multi-turn text interaction, camera navigation, combined text and camera control, and spatial memory. The model’s 54 dual-stream transformer blocks operate on DCAE-V latents, which compress video spatially and temporally. Standard inference uses the distilled model, a compatible decoder, and the Qwen text encoder; a lower-memory configuration targets one RTX 4090 with INT8 transformer weights, an FP8 decoder, CPU offloading, and a bounded history cache. The Apache-2.0 release also provides optional cinematic and robotics LoRA adapters. Its interactive demo uses a separate prompt-enhancement service, while offline use accepts prepared prompts.