Tencent's Hunyuan team has open-sourced Hy-Nomas, a video world model that generates a scene one segment at a time and lets users keep steering it with free-form text instructions and camera trajectories. It is built for researchers and developers who want to explore, redirect, and extend a generated world across turns, with a persistent key-value cache carrying earlier context forward so objects, environmental details, and spatial layout can stay consistent. The release matters now because the weights, inference code, training and annotation tools, a web demo, and a public benchmark arrived together, while the technical report is still marked as coming soon. The distilled setup denoises each segment in eight steps and writes 720 by 1280 video at 24 frames per second. The project site lists an 8 billion parameter model, continuous generation of up to 60 seconds, and a 360,000-token context. A low-memory configuration stores the diffusion transformer weights and cache in INT8, uses an FP8 decoder, and offloads stages to CPU so inference can run on a single RTX 4090. A cinematic LoRA and a robotics LoRA trained on the RealSource-World dual-arm dataset ship with the weights, and the repository includes the captioning and merge path for adapting the model to a custom video set. Hunyuan also published Hy-Nomas-Bench, four benchmarks with 280 test cases covering multi-turn text interaction, camera navigation, interleaved text and camera control, and spatial memory. Demo and gallery samples span real-world footage, CG, video game, and animation, in first-person and third-person views. The project is licensed under Apache 2.0.