Tencent Hunyuan published a technical report for Hy-Nomas and released the Hy-Nomas-Bench on 27 October.
The company previously shared information regarding the project on 21 October. Hy-Nomas functions as a video world model featuring an 8B diffusion backbone. Users can direct the model over multiple turns using continuous camera control, free-form text instructions, or a combination of both.
The system generates world segments one by one, conditioning each new part on the current input and a cached history. According to the report, the central design aligns structured video captions with action-aware causal attention. A vision-language prompt enhancer converts user requests and interaction history into structured conditions for the subsequent segment.
Training for the model follows a multi-stage curriculum. The distilled version of the model operates in four to eight denoising steps. The team claimed that Hy-Nomas outperformed existing world models.
Hy-Nomas-Bench includes 280 test cases covering four specific tasks: spatial memory, camera navigation, interleaved text and camera control, and multi-turn free-form text interaction. Tencent released both the model and the benchmarking tool under the Apache 2.0 license. Tencent provides an INT8 configuration that can run on a single RTX 4090.
Released adapters include a robotics LoRA trained on real dual-arm robot data and a cinematic LoRA. Outside developers have already added support for the model in WanGP, ComfyUI, and SGLang. One community demo reported real-time 720p generation on Google TPUs.
