Tencent opens the Hy-Nomas world model
On 21 October 2026 Tencent's Hunyuan team released the code and weights of Hy-Nomas, a video world model with an 8-billion-parameter diffusion backbone that generates an interactive world segment by segment from free-form text instructions and camera paths, under Apache 2.0. A technical report followed on arXiv on 27 October. One configuration runs on a single GeForce RTX 4090.
Why it matters
In the week DeepMind announced Genie 4 as a closed research preview, a comparable kind of model, one a user steers turn by turn, shipped with weights, training code, a benchmark and a recipe for running on one consumer graphics card. Whether it matches closed systems is not shown: the quality claim rests on the authors' own human evaluation.
Dates. The GitHub README (Tencent-Hunyuan/Hy-Nomas) lists among its news: 21 October 2026, "Initial code release: Training and inference code include LoRA fine-tuning and single-GPU inference. Get the model weights to start exploring.", with the interactive demo; 23 October a Diffusers inference example; 25 and 26 October community integrations (WanGP, SGLang, ComfyUI); and 27 October the technical report on arXiv. The Hugging Face repository was created on 21 October and the first commit of its present history is on 23 October. Hugging Face's paper page shows 26 October, the US Pacific date of the arXiv submission, which was at 06:48 UTC on 27 October. What the pages say. The model combines a segment-wise autoregressive diffusion transformer, a Qwen text encoder, camera-conditioning modules and a PixelUMM video decoder, with a persistent key/value cache carrying context between segments. The distilled configuration uses 8 denoising steps and outputs 720 × 1280 video at 24 frames per second; a separate configuration with INT8 weights, an INT8 cache, an FP8 decoder and CPU offload runs on one RTX 4090. A prompt-enhancement model, Nomas-PE, turns user instructions into structured prompts for the web demo and is a separate service. Hy-Nomas-Bench, a suite of four benchmarks with 280 test cases, is released as a dataset. Licence: Apache-2.0, except for third-party components, which keep their own licences. What the record does not claim. The abstract says "In human evaluation, Hy-Nomas shows superior performance over existing world models"; no figure from that evaluation is repeated and none was independently checked. No source read ties this release to Genie 4.