Tencent published Hy-Nomas on October 23, 2026, releasing an Apache-2.0 video world-generation model with open weights, multi-turn text interaction, camera control, and a documented low-memory path for video inference on one NVIDIA GeForce RTX 4090.
The Hy-Nomas demo and model page show the tradeoffs behind that setup. Tencent’s standard inference path uses eight GPUs in one inference group, while the one-card mode relies on INT8 DiT weights, an INT8 KV cache, an FP8 decoder, CPU offload, and up to 18 historical latent frames kept on GPU before older history is archived to system memory.
Tencent says Hy-Nomas generates video segment by segment and carries context forward through a persistent KV cache. The published defaults are 720×1280 output, 24 FPS playback, and eight denoising steps. Tencent also notes that 24 FPS refers to playback rate, not generation speed.
Alongside the release, Tencent published the 280-case Hy-Nomas-Bench for multi-turn text interaction, camera navigation, interleaved control, and spatial memory.