GitHub Repo
Tencent Open-Sources Hy-Nomas Interactive Video World Model, with a Single RTX 4090 Inference Setup
Tencent released Hy-Nomas weights on October 23, allowing users to steer video generation segment by segment with text and camera trajectories. The project also provides a low-memory inference setup and evaluation data, though long-range consistency and real-world latency still require independent validation.
Tencent Hunyuan has released Hy-Nomas, an interactive video world model. Hugging Face commit records show that the weights were first published on October 23 under the Apache-2.0 license. The accompanying GitHub project provides code for training, inference, data annotation, and an interactive interface. Before generating the next video segment, users can modify text instructions or camera trajectories to gradually change the scene, objects, and actions. Weight commit history, official repository
Architecturally, Hy-Nomas runs diffusion in a compressed video latent space and generates video autoregressively, one segment at a time. The text encoder, camera conditioning, and historical KV cache all influence each new segment. The distilled configuration uses eight denoising steps, followed by decoding with a fine-tuned PixelUMM. The official output specifications are 720×1280 at 24 frames per second. These are video playback specifications and do not establish that the model can generate video in real time. Official model card
A key deployment trade-off is between historical context caching and video memory usage. The standard configuration retains the full history. The single RTX 4090 configuration combines INT8 weights and KV cache, an FP8 decoder, and CPU offloading, while limiting the historical content used by GPU attention. The official documentation warns that offloading requires sufficient system memory and adds data transfer costs between the CPU and GPU. Engineering teams should therefore measure peak memory, latency, and scene consistency at their target resolution and segment length. Deployment documentation
The release also includes Hy-Nomas-Bench, a benchmark with 280 cases covering multi-turn text interaction, camera navigation, interleaved text and camera control, and spatial memory. It provides structured prompts, interaction sequences, and reference trajectories as a starting point for reproducible testing. However, the public test materials alone do not establish that the model outperforms other systems. Official evaluation dataset
The release lets researchers inspect the implementation of segment-by-segment generation and historical attention directly, and fine-tune the model on their own videos using LoRA. Areas to watch include the technical report, reproducible comparison results, and how quantization and historical context cropping affect long-range consistency. As of the time of review, the repository still lists the technical report as forthcoming; the reliability of the model’s world memory remains to be tested. Project overview