Tencent Hunyuan publishes Hy-Nomas report, releases 280-case world model benchmark under Apache 2.0

Tencent Hunyuan has released Hy-Nomas-Bench, a 280-case benchmark for interactive video world models, alongside the technical report for Hy-Nomas, its open interactive world generation model. The report was posted to arXiv on October 27. Code and weights for the model were first released on October 21.

The model

Hy-Nomas is a video world model with an 8B diffusion backbone. Users steer it over multiple turns with free-form text instructions, continuous camera control, or both. It generates the world segment by segment, conditioning each new segment on the current input and on a cached history, so scenes stay consistent over time.

According to the report, the central design aligns structured video captions with action-aware causal attention. Descriptive context stays available across segments, while action instructions apply only to the time window they target. A vision-language prompt enhancer converts user requests and the interaction history into structured conditions for the next segment.

Training

Training follows a multi-stage curriculum:

1. Causal generation is introduced first.
2. Text interaction and camera control come next.
3. The model then trains on histories it generated itself.
4. A final diffusion distillation step produces the released fast model.

The distilled model runs in four to eight denoising steps. The team says Hy-Nomas outperformed existing world models in human evaluation.

Hy-Nomas-Bench

The benchmark contains 280 test cases across four tasks:

• Multi-turn free-form text interaction
• Camera navigation
• Interleaved text and camera control
• Spatial memory

It checks whether a model follows successive instructions, produces physically plausible responses, executes camera moves accurately, and keeps a scene consistent when the camera returns to a previously seen place. The suite includes prompts, structured captions, interaction sequences and reference camera trajectories. It is released under Apache 2.0.

Release contents and licensing

The model is also licensed under Apache 2.0. Tencent provides:

• An INT8 configuration that runs on a single RTX 4090, using sliding-window history and CPU offload
• A Diffusers pipeline
• LoRA fine-tuning code
• An interactive web demo
• Two released adapters: a cinematic LoRA and a robotics LoRA trained on real dual-arm robot data

Outside developers have already added support in SGLang, ComfyUI and WanGP. One community demo reports real-time 720p generation on Google TPUs.

Context window, pricing and numeric benchmark scores are not applicable or not yet disclosed in the material published so far.

What this means

The benchmark is the main new artifact here. Hy-Nomas-Bench tests four things at once: instruction following over successive turns, camera accuracy, physical plausibility and memory of revisited places. Because the prompts, structured captions, interaction sequences and reference camera trajectories ship with an Apache 2.0 license, other teams can run the same cases against their own world models.

The human evaluation result is Tencent's own claim, and the report's reference to existing world models should be read in that light until others reproduce the comparison on the released suite.

The INT8 single-RTX 4090 configuration, the Diffusers pipeline and the LoRA code mean the model can be run and adapted on one consumer GPU. The robotics LoRA, trained on real dual-arm robot data, shows the model being adapted beyond cinematic video generation.