Hy-Nomas
Released Oct 21, 2026
Tencent Hunyuan's open-weight free-form interactive world model that builds a world one video segment at a time from text, which users keep steering with multi-turn text instructions and camera trajectories as generation continues.
- Output
- Live video stream
- Interactive
- —
- Inputs
- text, actions
- Max session
- —
- Resolution
- 1280x720
- Frame rate
- 24 fps
- Access
- Generally available
- Licence
- Apache License 2.0
Strengths & weaknesses
Strengths
- Multi-turn steering mid-generation: add objects, change style or scene, describe actions, move the camera
- 720x1280 output at 24 FPS from an 8-step distilled model
- Runs on a single RTX 4090 via a low-memory INT8/FP8 configuration with CPU offload
- Apache-2.0 weights with inference and LoRA training code, plus Hy-Nomas-Bench (280 cases)
Weaknesses
- No official generation-speed figure; Tencent says speed depends on hardware, resolution, segment length and history
- Text and camera control only; no image conditioning documented
releaseDate = GitHub README 'initial code release' 2026-10-21 (repo created that day; comfyui-wiki gives the same date). fps 24 is the output playback rate, not generation speed; a community ComfyUI port reports slower than real time on an RTX 4090 (~3.5-4 s per 1 s turn at 832x480), which is why interactive is null rather than true. 'actions' = camera trajectories / keyboard navigation. commercialUse true applies to the Apache-2.0 release only: the card says third-party components keep their own licences, and community repackagers (nynxz, DeepBeepMeep) say the bundled PixelUMM decoder is a fine-tune under NVIDIA's One-Way Noncommercial License. That decoder claim isn't confirmed on Tencent's own pages. Weights include Cinematic and Robotics LoRAs and a Nomas-PE prompt enhancer. Separate line from HY-World 1.5 (WorldPlay) and HY-World 2.0.
Links
API model strings
- huggingface
tencent/Hy-Nomas - huggingface
tencent/Hy-Nomas-Diffusers