Tencent Hunyuan releases Hy-Nomas, an open-source interactive world generation model
Tencent Hunyuan released Hy-Nomas, a free-form interactive world generation model controlled by multi-turn text and camera inputs. It runs on a single RTX 4090 with low-memory inference and supports LoRA fine-tuning. The initial code release was on Oct 21, 2026.
China context
- Original name
- 混元
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can access the model weights and code from GitHub and Hugging Face, and run inference on a single RTX 4090. LoRA fine-tuning enables customization for specific world generation tasks.
- For investors
- Tencent's release of an open-weights world generation model signals its commitment to the open-source AI ecosystem and may attract developer mindshare, potentially influencing adoption of Tencent's broader AI offerings.
Hy-Nomas is a free-form interactive video world generation model that supports multi-turn text prompt control and camera control. The repository provides training, inference, interactive demo, and data annotation code. Key features include segment-based generation with lifelong KV cache, free-form prompt control, camera control, distilled inference using 4-8 denoising steps and a finetuned PixelUMM decoder for 32-channel DCAE-V latents, and an interactive web demo. Community integrations include SGLang, ComfyUI, Reactor Runtime, WanGP, and Google TPU demos.
Hy-Nomas uses autoregressive segment-wise generation with a lifelong KV cache for consistency, conditioned on segment-level prompts or camera trajectories. Inference is distilled to 4-8 denoising steps with a finetuned PixelUMM decoder for 32-channel DCAE-V latents, enabling single RTX 4090 low-memory inference. LoRA fine-tuning is supported for downstream tasks.
Developers outside China can now access a Tencent open-weights model for interactive world generation that runs on a single consumer GPU, reducing the compute barrier for building world-model applications. This puts pressure on closed world-generation APIs by offering a self-hostable alternative with LoRA fine-tuning.
Hy-Nomas provides a low-cost, open-weights option for developers building interactive world generation applications, with single-GPU inference and LoRA fine-tuning enabling customization. The interactive demo and community integrations lower the barrier to experimentation.
A measurable signal is the number of community integrations and downstream projects that adopt Hy-Nomas, as well as whether Tencent releases benchmark results or a technical report with quantitative evaluations. The availability of the Hy-Nomas-Bench dataset on Hugging Face may enable independent benchmarking.