Tencent Hunyuan Open-Sources Hy-Nomas, an Interactive World Model

Tencent’s Hunyuan team has released Hy-Nomas, an AI world model that generates video one segment at a time and lets users change the scene with text or steer the camera as it goes. The model weights are on Hugging Face under the Apache 2.0 licence, with inference code on GitHub, a test set and a public demo.

The release is light on evidence of how well it works. Tencent has published a benchmark of 280 test cases but no scores, and the technical report is marked “Coming Soon”. As of Friday afternoon there was no formal announcement beyond the repositories and the demo site.

What Hy-Nomas does

The model card describes Hy-Nomas as “a video world generation model that supports free-form, multi-turn interaction through text instructions and camera control. It generates a world one video segment at a time, allowing users to introduce objects, change styles and scenes, describe actions, or control the camera as generation continues.”

In practice, a user describes an opening scene, then adds instructions for each following segment, such as a new object, a change of style or an action. Camera trajectories can be supplied alongside the text to move the viewpoint. Tencent’s own example starts with a fox on a forest path and then has glowing flowers open beside it as it walks on.

The demo site lets visitors start a world, continue it with text or camera paths and download the result. The demo relies on a separate prompt-enhancement service, Nomas-PE, which turns plain instructions into the structured prompts the model expects.

How it is built

Hy-Nomas is a segment-autoregressive diffusion transformer. Each new segment is generated by a diffusion process, while the model moves forward one segment at a time and carries earlier context in a persistent key/value cache. The card lists 54 dual-stream transformer blocks, a text encoder taken from Alibaba’s Qwen2.5-VL-7B-Instruct model, and camera information fed in through cross-attention. A fine-tuned decoder called PixelUMM turns the compressed output into video.

The distilled version uses eight denoising steps per segment and outputs 720 × 1280 video at 24 frames per second. Tencent notes that 24 FPS is the playback rate, not a generation speed: speed “depends on hardware, resolution, segment length, history, and the selected configuration”.

The standard setup splits work across eight GPUs. A second configuration targets a single Nvidia GeForce RTX 4090, storing the model weights and cache in compressed INT8 form, using an FP8 decoder and moving components on and off the GPU as needed. The release also includes an optional Cinematic adapter for visual quality and a Robotics adapter trained on a public dataset of real-world robotic-arm footage, which Tencent says improves the robotics simulation capability of generated worlds.

Hy-Nomas should not be confused with HY-World 2.0, a separate Tencent Hunyuan project released in April that produces 3D world meshes and Gaussian splats rather than streaming video. Tencent’s team has also recently open-sourced Hy-Memory 2.0, a memory toolkit for AI agents.

The benchmark ships without results

Alongside the model, Tencent released Hy-Nomas-Bench, four test sets with 280 cases in total. They cover multi-turn text instructions, camera navigation, mixed text and camera control, and spatial memory, meaning whether a model shows the same place consistently when the camera returns to it. The dataset includes structured prompts, interaction sequences and 220 reference camera trajectories.

What it does not include is any score for Hy-Nomas or a rival. The model card says plainly that its architecture and sampling defaults “are configuration details, rather than benchmark results”.

Open weights and a test set make Hy-Nomas checkable by outsiders

The most useful part of the release may be that others can verify it. Several recent world models have arrived with headline numbers that are hard to test: TSN noted that Odyssey’s physics score for Odyssey-3 was a best-of-eight, self-reported result. Hy-Nomas reverses the order. It offers no claims, but it publishes the weights, the code and the test cases under a licence that allows commercial use.

That means independent researchers can run the 280 cases themselves and report how well the model follows instructions over many turns and whether it remembers a scene. Tencent’s eventual technical report will be easier to judge if outside results already exist. The single-GPU configuration matters here too. A model that needs an eight-GPU server is open in name to most labs; one that runs on a high-end consumer card can actually be tested by universities and small teams, even if it is slow.

The Apache 2.0 licence also covers the benchmark, so other world-model developers can adopt it. Interactive world models still lack shared tests, and a public, well-documented set aimed at multi-turn control and spatial memory could prove more lasting than the model itself.

Memory over long sessions is the limit Tencent itself flags

The card is candid about one constraint. In the standard setup, the model keeps its full history, and that “consumes increasing cache memory” as generation continues. The single-GPU version caps how much history sits on the graphics card at 18 latent frames and archives the rest in system memory, which still keeps growing.

This is the central tension for interactive world models. Remembering more of what has already been generated is what keeps a world consistent when a user turns back to a place they have seen. Remembering everything becomes expensive. Other teams are tackling the same trade-off: TSN has covered Helix World, a real-time world model from Noiz AI and HKUST, and investors are backing real-time world-model start-ups such as Reactor.

Hy-Nomas-Bench includes a spatial-memory test precisely to measure this, but until results appear there is no way to tell how well the bounded single-GPU history holds up against the full eight-GPU version. The Robotics adapter points to one reason the question matters: world models used to simulate robot tasks need scenes that stay physically and spatially consistent.

What happens next

The next milestone is Tencent’s technical report, which the GitHub page still lists as “Coming Soon”, and with it the first Hy-Nomas scores on its own benchmark. Independent runs of the 280 test cases, especially the spatial-memory set on a single RTX 4090, would give the first outside measure of what this open world model can do.

Sources

• tencent/Hy-Nomas model card, Hugging Face, October 2026 (Apache 2.0)
• Tencent-Hunyuan/Hy-Nomas repository, GitHub, October 2026
• tencent/Hy-Nomas-Bench dataset, Hugging Face, October 2026
• Hy-Nomas demo and project site, Tencent Hunyuan
• HY-World 2.0 repository, GitHub (context, separate project)

Related stories

• Tencent Hunyuan Open-Sources Hy-Memory 2.0 Memory SDK and Writer Model
• Noiz AI and HKUST Release Helix World, a Real-Time World Model Whose Sound Turns With the Camera
• Odyssey Opens Its Odyssey-3 World Model to the Public; Its Headline Physics Score Is a Best-of-Eight, and Every Benchmark Is Odyssey’s Own
• Reactor’s Series A: NVentures and Sapphire join the real-time world-model bet