The Pulse
Tencent Opens Hy-Nomas, an Interactive Video-World Model
Tencent has released Hy-Nomas, a video-generation model users can steer with text prompts and camera paths as it creates successive segments. The release includes model weights, code and a 280-case benchmark, but the benchmark package conta
AI.info Team · Written with AI · fact-checked ·
Tencent’s Hy-Nomas promises a world users can keep changing as it generates video, but the release arrives with a notable gap between capability and evidence: its benchmark has 280 test cases and no published evaluation scores. Released on October 23, the model takes text instructions or camera paths between video segments, while Tencent’s technical report is still marked “Coming Soon.”
Hy-Nomas changes the scene between segments
Hy-Nomas generates video in successive segments rather than treating a prompt as a one-time instruction for a complete clip. Users can add objects, change a scene or its style, describe actions and control the camera as generation continues. A persistent key-value cache carries context from earlier segments into later ones, the mechanism Tencent uses to maintain continuity.
“Hy-Nomas is a video world generation model that supports free-form, multi-turn interaction through text instructions and camera control,” Tencent says in its model card. The release provides code for training, inference, an interactive demo and data annotation, alongside weights hosted on Hugging Face. The project is licensed under Apache 2.0; the model card cautions that third-party components remain under their own licenses.
The “world” in the name describes the ongoing video, not an exported, editable 3D environment. Tencent’s documentation specifies video output and camera control; it does not describe a 3D scene file or a game engine. That distinction matters for developers assessing whether the model fits video production, interactive prototypes or applications requiring structured 3D assets.
One GPU is possible, with trade-offs
The standard inference configuration uses eight GPUs in one sequence-parallel group. Tencent also provides a low-memory configuration for a single NVIDIA GeForce RTX 4090, using INT8 weights and cache, an FP8 decoder and CPU offload. The single-GPU setup keeps up to 18 historical latent frames available on the GPU and archives earlier history on the CPU.
That configuration has costs: Tencent says the offloaded models and growing history archive require enough system RAM, while transfers between CPU and GPU add latency. Peak GPU memory also varies with resolution and segment length. The standard distilled setup lists 720 × 1,280 output at 24 frames per second with eight denoising steps, but 24 FPS is the video’s playback rate, not a claim that generation runs in real time.
The 280-case benchmark has no scores
Tencent released Hy-Nomas-Bench with four test suites: 60 cases for multi-turn text control, 100 for camera navigation, 60 for interleaved text and camera control, and 60 for spatial memory. The package includes 220 reference camera trajectories and test prompts covering real-world, game, computer-generated and animated environments. Its cases request videos ranging from 8 to 86 seconds.
The benchmark release supplies test inputs and instructions, but not generated videos, reference or revisit frames, or evaluation scores. Developers can run the tests using the released model and code, but the package itself does not show how Hy-Nomas performs or provide a scored comparison with other systems. That leaves the model’s stated interaction and memory features without published benchmark results in this release.
Demo and technical report remain the next checks
Tencent’s online demo lets users begin a world, continue it with text instructions or camera paths, and download the resulting video. For local use, the repository includes inference instructions and LoRA fine-tuning support for custom video datasets. The company also lists cinematic and robotics adapters among the model assets.
The available materials establish what developers can download and how the model is configured; they do not independently verify output quality or generation speed. The benchmark’s missing scores and the forthcoming technical report are the clearest limits on evaluating the release today. For now, Hy-Nomas is an open video-generation system with interactive controls, an accessible demo and a test suite users can run themselves.