Tencent's Hy-Nomas world model ships with no published scores

Tencent Hunyuan pushed Hy-Nomas, an open interactive video world model, live on October 23, 2026, and published no results from it. The Apache-2.0 weights are on Hugging Face, the training and inference code is on GitHub, and a four-part benchmark shipped beside them. What the release does not include is a single score, and the "Technical Report (Coming Soon)" badge on the model card links to a file that returns a 404.

Hy-Nomas is a segment-autoregressive diffusion transformer. It builds a world one segment at a time, and between segments you can hand it a new free-form text instruction or a camera trajectory. A persistent key/value cache carries the generated history forward, so objects and layouts from earlier segments stay in the world. The card pairs a Qwen2.5-VL-7B text encoder with camera-conditioning modules and a fine-tuned PixelUMM decoder.

The distilled inference configuration outputs 720 x 1280 video at 24 FPS in 8 denoising steps. The card is explicit that 24 FPS is the playback rate of the output, not a generation speed, and Tencent publishes no latency figure at all. Camera control is its own conditioning path rather than prose: a trajectory file holds camera-to-world matrices and intrinsics, and the model attends to them in every block.

Two LoRA adapters ship with the weights, one Cinematic and one for dual-arm robotics trained on a real-world robotic-arm dataset. The default setup is heavy, using 8-way sequence parallelism and eight GPUs; the training recipe in the code repository runs on 128.

A second configuration targets a single GeForce RTX 4090. It stores the DiT weights and the key/value cache in INT8, decodes in FP8, moves the encoder and both networks between CPU and GPU for their own stages, and caps active history at 18 latent frames. Tencent warns that the CPU offload needs enough system RAM and adds transfer overhead.

Hy-Nomas-Bench covers multi-turn text control, camera navigation, interleaved text and camera control, and spatial memory across 280 test cases. The project site shows 83 generated worlds and lists 8B parameters, a 720p output and a 60-second maximum duration. The only announcement is a post on X from the official Tencent Hunyuan account, which by October 24 had about 40,700 views.

The hosted demo is open now.

Try Hy-Nomas in a browser at hy-nomas.hy.tencent.com; self-hosting starts from the hynomasoss/hynomasoss:cu130 Docker image in the repository and needs CUDA 13.