Tencent open-sources Hy-Nomas video world model you can direct mid-generation, runs on one RTX 4090
lmoroney · x · 2026-10-25
Tencent Hunyuan open-sourced Hy-Nomas, a video world model you steer as it generates: it builds scenes segment-by-segment autoregressively, and between segments you can add characters, change style, describe actions, or set camera paths, with a persistent KV cache keeping the world consistent. Stack: segment-autoregressive DiT + Qwen text encoder + camera conditioning + PixelUMM decoder, distilled to 8 denoising steps at 720×1280/24FPS. Weights are Apache 2.0 on Hugging Face; standard config needs 8 GPUs, but an INT8 + CPU offload config runs on a single RTX 4090. Ships with a cinematic LoRA and Hy-Nomas-Bench (280 test cases). Great as a previs sandbox: feed scenes as 5-second beats one at a time.
More from Embodied
- 3D Printing New Shoes for a Home Robot — JohnVial · 2026-10-26
- Apple's iPad mini gets OLED, 2nm chip, 32-core Neural Engine for $699 — shauntrennery · 2026-10-26
- StackChan desk robot gains multilingual voice and can now act from scripts — alexcovo_eth · 2026-10-26
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Shenzhen company combines humanoid robots with industrial robot arms — RoboBalaji · 2026-10-26