Research into systems that model the dynamics of environments, rather than only producing fixed sequences of imagery, has expanded as interest grows in supporting longer interactions, agent training, and flexible simulation.
Static video generators have improved in visual quality and length, yet many still require a complete description in advance and discard context between separate generations.
In this setting, Tencent's Hunyuan team has released 'Hy-Nomas,' an open-source interactive video world model that produces scenes in successive segments while accepting free-form text prompts and camera trajectories as ongoing control inputs.
The system relies on a Diffusion Transformer architecture and generates video autoregressively.
Each new segment is conditioned on previously produced content retained in a lifelong key-value cache, an approach intended to maintain object placement, spatial relationships, and scene details across multiple turns.
Users can interleave natural-language instructions, such as altering an object, changing lighting conditions, or directing an action, with explicit camera paths inside the same session.
Reported output specifications include 720p resolution at 24 frames per second.
Distillation reduces the number of denoising steps required for inference to 8, and the release includes an INT8 quantized version of the primary weights along with an FP8 decoder.
These optimizations, combined with sliding-window history and CPU offloading, allow the model to run on a single RTX 4090.
A technical report has been listed as forthcoming.
The published materials also contain training and data-annotation code, LoRA adapters oriented toward cinematic and dual-arm robotics domains, and a benchmark consisting of roughly 280 cases covering multi-turn text interaction, camera navigation, interleaved control, and spatial memory.
An online demonstration and options for local deployment are available under an Apache 2.0 license.
Public demonstration material illustrates sequences that include an underwater passage containing a manta ray and stone arches, a paraglider descending over snow-covered terrain, a motorboat crossing open water under keyboard-style camera input, a cyclist traveling along a palm-lined street for more than half a minute, and a dual robotic arm manipulating objects on a tabletop.
Shortly after the initial release, community ports for tools such as ComfyUI appeared.
Early user reports have noted both the responsiveness of camera controls and the practical memory requirements once quantization and offloading are applied.
The model is positioned as generating continuous video under user direction rather than a single fixed clip, with the cache providing a form of world memory across segments.
These properties place Hy-Nomas among several recent interactive world models that differ in architecture, output representation, sensory scope, and access model.
Among others include Google DeepMind's Genie 3. The model is able to generate real-time, navigable environments with multi-minute consistency and mid-session event prompting. Then, World Labs' Atlas, which can processes text, images, video, and 3D data to generate up to one minute of camera-controlled 1440p video and explicit 3D representations like Gaussian splats, currently available through early access for select partners. Helix World, created by Noiz AI and HKUST, synthesizes synced video and spatial stereo audio at 768 by 512 at 24 frames per second based on camera and keyboard inputs, with its code and a preview checkpoint publicly released. SpAItial AI's Echo-2 creates persistent, edit-friendly 3D Gaussian splat scenes directly from text or image inputs, offering real-time rendering and object manipulation through a browser application. Runway's GWM Worlds 2 provides continuous 720p interactive video and 48 kHz audio driven by timestamped text actions, supporting unbounded session lengths in a research preview available by request.
Across these approaches, differences appear in whether the output is sequential video or an explicit three-dimensional representation, whether audio is generated jointly with vision, the degree of open access to weights and code, and the emphasis placed on real-time streaming versus segment-wise generation with persistent memory.
However, the system is not without flaws. Issues such as visual artifacts during sharp camera maneuvers, VRAM-intensive local execution, and temporal bleeding over long sessions highlight that the technology is still in an exploratory phase rather than ready for drop-in commercial engine replacement.
Ultimately, Hy-Nomas offers a functional framework for real-time, promptable spatial environments, representing a useful step forward for open-source world modeling. It demonstrates how interactive video generation can react dynamically to user actions rather than following fixed paths.
