Tencent's Hunyuan team has released Hy-Nomas, an open model that generates a video world piece by piece and lets a user steer it as it unfolds, with a new text instruction or a camera movement at each turn. The code appeared on GitHub on October 21, 2026, the weights were uploaded to Hugging Face the same day, and the team announced the release from its official X account, Tencent Hy, on October 23. Code and weights are published under the Apache 2.0 licence.
Unlike a text-to-video system that renders one fixed clip from a single prompt, Hy-Nomas works in segments. Each segment is produced by a diffusion transformer conditioned on the latest instruction, an optional camera trajectory and a key-value cache that carries what has been generated so far. Users can add objects, change the style or scene, describe actions or move the viewpoint, and the model continues the same world. Tencent's project page lists 8 billion parameters, 720p output, up to 60 seconds of continuous video and a context of 360,000 tokens, and shows demos in which a large camera move returns to an earlier view with the background preserved.
According to the project page and the model card, the backbone has 54 dual-stream MMDiT blocks linking text and video, with camera cross-attention guiding the video stream. A Qwen2.5-VL text encoder reads the instructions, and a fine-tuned PixelUMM decoder turns 32-channel DCAE-V latents into frames. Action-aware attention keeps scene descriptions visible across segments while restricting action text to its own time window, and a training method Tencent calls Segment Forcing exposes the model to its own predicted history. A distilled checkpoint generates each segment in four to eight denoising steps.
The release is unusually complete for a world model. The Hugging Face repository holds a base checkpoint for fine-tuning, the distilled model and an INT8 version, LoRA adapters for cinematic footage and for dual-arm robotics trained on the RealSource-World dataset, the video autoencoder and decoders, and a prompt-enhancement model called Nomas-PE that rewrites user instructions for the demo. The GitHub repository adds training, inference and data-annotation code, a web demo and a streaming integration with Reactor Runtime. Tencent also published Hy-Nomas-Bench, 280 test cases covering multi-turn text control, camera control, interleaved control and spatial memory, plus a Diffusers pipeline that for now installs from a fork rather than the main library.
Hardware needs vary widely. Tencent says the model can run on a single NVIDIA GeForce RTX 4090 by storing weights and the cache in INT8, bounding the history with a sliding window and offloading components to the CPU, which shortens the model's memory and slows generation. The tested real-time streaming profile, by contrast, uses eight NVIDIA B200 GPUs, at least 128 GB of host RAM and about 80 GB of checkpoints, and the example fine-tuning run is sized for 128 GPUs.
Early uptake is visible but still modest. By October 26 the announcement post had about 60,000 views and 900 likes on X, community ports for ComfyUI had appeared on Hugging Face within hours, and the GitHub repository had about 186 stars. A technical report is marked as coming soon, and Tencent has not yet published quantitative comparisons with other world models, so claims about quality and consistency rest for now on its own demos.
For developers working on games, simulation and embodied AI, an openly licensed interactive world model that ships with fine-tuning code and a robotics adapter is a notable addition to a field where many of the best-known systems cannot be downloaded. How well it holds up will become clearer once the report and independent tests arrive.