zhineng.dev
English
Get started
← All models

Hy-Nomas

Published Publisher Tencent

Overview

Hy-Nomas is an open-weight interactive video world-generation model. It generates a world in successive video segments, guided by multi-turn text instructions and camera movement. Its diffusion backbone has approximately 8 billion parameters. (github.com)

Capabilities and uses

  • Generates scenes, objects, styles and actions from free-form prompts, with camera trajectories as an additional control. It uses cached history to support continuity across segments. (github.com)
  • The interactive demo supports multiple turns; image-reference initialization was added on October 31, 2026. (github.com)
  • Supports LoRA fine-tuning on custom video datasets. Tencent provides cinematic and robotics adapters, making interactive content creation, research and robotics-simulation experiments potential uses. (github.com)

Availability and limitations

Tencent provides model weights, training and inference code, and an interactive demo. The project is licensed under Apache 2.0. (github.com)

A low-memory configuration can run on one NVIDIA RTX 4090, but uses quantization, a bounded sliding-window history and CPU offloading. It requires sufficient system RAM; transfer overhead and peak GPU memory depend on the setup, resolution and segment length. Physical consistency should not be assumed: the technical report describes robust physical dynamics in video-based world models as an open research question. (github.com)

Release date

Initial code and interactive demo: October 21, 2026. (github.com)

Pricing

As of November 15, 2026, Hy-Nomas-specific official input and output pricing could not be verified in the Tencent Cloud mainland China and international pricing pages checked. No CNY or USD amount or billing unit is asserted. Self-hosting entails infrastructure costs; prices for other Hunyuan models should not be treated as Hy-Nomas prices. (cloud.tencent.com)

Official sources