Hy-Nomas: Towards Free-form Interactive World Generation
modelYour notes
Tencent Hunyuan's interactive video world model, released with weights under Apache 2.0 on October 21, 2026 (paper October 27). It generates a world one video segment at a time, each conditioned on a free-form text instruction, a camera trajectory and the cached history of earlier segments. A user can add objects, change style or scene, describe actions or move the camera as generation continues. The backbone is an 8B dense MMDiT (54 double-stream blocks, hidden size 2,048) that runs diffusion within a segment and autoregression across segments. A persistent KV cache carries context forward, and attention is causal across segments. Text comes in through the language model of Qwen2.5-VL-7B, and camera extrinsics through cross-attention.
The paper's central design pairs structured video captions with action-aware causal attention: descriptive context stays visible to every later segment, while an action instruction is confined to the time window it was issued for. A separate vision-language Prompt Enhancer rewrites user requests and the interaction history into structured conditions for the next segment. Training follows a curriculum that adds causal generation, text interaction, camera control and then training on the model's own generated histories, followed by diffusion distillation to 8 denoising steps at 720×1280 and 24 fps. A bounded-history configuration runs on a single RTX 4090. Tencent also released Hy-Nomas-Bench, 280 cases across multi-turn text interaction, camera navigation, interleaved control and spatial memory. Its claim of beating existing world models rests on its own human evaluation.