Tencent released Hy-Nomas, an open-source video world model that builds a scene one video segment at a time while you direct each new segment with text instructions, camera trajectories, or both, according to the Hy-Nomas project page. That could make it useful for previs and iterative scene exploration, since you can keep adjusting subjects, actions, style, or camera movement across turns.
- Each turn can redirect the world. Tencent says you can introduce objects, change styles and scenes, describe actions, or control the camera as generation continues.
- Tencent lists 720p output and up to 60 seconds. The project page specifies 8B parameters, 720p, a 60-second maximum duration, and 360K-token context; the model card lists 720 x 1280 at 24 FPS for standard inference.
- Code and weights ship under Apache-2.0. The repository includes training, inference, interactive demo, and data annotation code, while third-party components keep their own licenses.
- Tencent says one RTX 4090 can run a low-memory video inference configuration. It relies on INT8 weights, an FP8 decoder, and CPU offload; the standard setup uses eight GPUs.
- Performance claims come from Tencent. Tencent had not published its technical report at release, and the benchmark released with the model is Tencent's own.
Text instructions and camera trajectories condition every new segment
Hy-Nomas combines a segment-autoregressive diffusion transformer, a Qwen text encoder, camera-conditioning modules, and a PixelUMM video decoder. A persistent key/value (KV) cache carries context from previous segments, so each continuation is conditioned on what the model already generated.
Text control is free-form. In Tencent's example prompt, the first segment describes a small orange fox on a stone path in a luminous forest, and the next segment uses a description_change field to have glowing blue flowers open beside the path as the fox keeps walking.
Camera control uses trajectory files containing camera-to-world matrices and camera intrinsics, aligned to the requested duration and frame rate. The interactive demo lets you start a world, continue it with text instructions or camera paths, and download the generated video.
Tencent's project page shows what it calls consistent world memory, including demonstrations of revisiting generated scenes. Those are Tencent's demonstrations, and we have not tested spatial or temporal consistency ourselves.
The output is generated video. Tencent does not describe reconstructing captured footage or exporting an editable 3D scene.
Default inference uses eight GPUs, while a single RTX 4090 needs memory trade-offs
The standard configuration runs the distilled model with 8 denoising steps and 8-way sequence parallelism, so its default launch requires eight GPUs in one inference group.
The low-memory path targets a single NVIDIA GeForce RTX 4090 using a quantized INT8 checkpoint and an FP8 decoder. It changes several things to fit:
- INT8 storage. DiT weights and the KV cache are stored in INT8, with BF16 computation retained.
- Bounded GPU history. The GPU keeps up to 18 historical latent frames, with older history archived on CPU.
- CPU offload. The text encoder, DiT, and decoder move onto the GPU only for their processing stages.
Tencent notes that system RAM must hold the offloaded models and the growing history archive, and that CPU-GPU transfers add latency. The 24 FPS figure is the output's playback rate; generation speed depends on hardware, resolution, segment length, history, and configuration.
The single-GPU claim covers video inference only. The web demo also requires Nomas-PE, a separate prompt-enhancement model that converts user instructions into structured prompts and runs on its own serving resources.
Cinematic and robotics LoRAs, a fine-tuning pipeline, and a vendor-built benchmark ship with the model
The release includes a Cinematic LoRA trained on cinematic data, which Tencent says enhances the visual quality of generated worlds when merged into the distilled model. A second adapter, trained on the RealSource-World robotic-arm dataset, targets robotics simulation.
The adapters are meant to be merged into the distilled floating-point checkpoint. The INT8 loader used for single-GPU inference does not support loading a separate LoRA adapter at inference time.
For custom adapters, the GitHub repository includes a captioning harness that prepares training data from local videos. Training starts from the non-distilled checkpoint, and the learned weights are then merged into the distilled checkpoint for inference.
Tencent's example training configuration runs 2,000 steps, with a sample 128-GPU launch that users can resize for their hardware.
Tencent also released Hy-Nomas-Bench, four benchmarks with 280 test cases covering multi-turn text interaction, camera navigation, interleaved text and camera control, and spatial memory. Because Tencent built and released it, it does not serve as independent validation.
Related world-model coverage: Skywork's Matrix-Game 3.0 world model also targets 720p interactive scene generation, in real time.
