What it is
Hy-Nomas is Tencent Hunyuan's interactive video world model, open-sourced on October 23, 2026 under Apache 2.0 (model weights, inference code, and eval set all open, no gate).
A "world model" here means: it generates video segment by segment while keeping the scene consistent as you interact — like a game engine, but the frames are generated, not rendered. Unlike Sora / Runway's "one prompt → one clip," Hy-Nomas is generate-while-interact: you use text or camera trajectories to change the scene, add objects, switch style, or swing the camera at each continuation.
Technically it's a segment-autoregressive diffusion transformer: each segment is denoised by diffusion, and across segments a "lifelong KV cache" remembers prior objects, lighting, and geometry, so looking back at the same place still lines up. Architecture: 54 dual-stream MMDiT blocks, hidden dim 2048, Qwen2.5-VL-7B-Instruct text encoder, video compressed in a 32-channel DCAE-V latent space (32× spatial, 4× temporal), finally decoded to pixels by the PixelUMM decoder. In the distilled config each segment is 8 denoising steps, output 720×1280 @ 24 FPS (24 FPS is playback rate, not generation speed).
Hands-on / install
Easiest path: just play the demo. The online demo at hy-nomas.hy.tencent.com lets you open a world, continue with text or camera trajectories, and download results. Note the demo depends on a separate prompt-enhancement service, Nomas-PE (not needed offline).
Run it yourself: two hardware routes
The standard route needs 8 GPUs (8-way sequence parallel). The real selling point is the single RTX 4090 route: store DiT weights and KV cache in INT8 while computing in BF16, decoder in FP8 (CUTLASS), shuttling text encoder / DiT / VAE stages between CPU and GPU, capping the history window at 18 latent frames with older frames archived to system RAM.
Mind the disk math: the standard demo path needs distilled DiT (34.2GB) + PixelUMM decoder (31.55GB) + DCAE-V (4.28GB) + Qwen encoder ~16GB — ~86GB before you can even generate; the single-card INT8 route compresses to ~38GB + encoder.
The environment is specific: Linux, Python 3.12+, CUDA 13.0 Toolkit, PyTorch 2.10.0, plus FlashAttention 2 / 3 and MagiAttention. The official build image hynomasoss/hynomasoss:cu130 is recommended — don't hand-compile the attention kernels.
Capability boundaries (read before you spend on a GPU)
- No public scores: the bundled Hy-Nomas-Bench has 280 cases across 4 suites (multi-turn text control, camera navigation, text+camera hybrid, spatial memory) with 220 reference camera trajectories, but not a single score is given. The model card itself states "architecture and sampling defaults are configuration details, not benchmark results." So right now you cannot place it numerically against any model — you have to run those 280 cases yourself.
- Technical report coming soon: comparison benchmarks are not out, and no inference provider hosts it (the HF page states "not deployed by any inference provider").
- LoRA can't stack with the INT8 single-card route: the Cinematic (6.12GB) and Robotics (12.24GB, trained on the RealSource-World dual-arm dataset) adapters are not supported for standalone LoRA under the INT8 loader — the two paths are incompatible.
- Not HY-World 2.0: don't confuse it with Tencent's April release — HY-World 2.0 outputs 3D meshes and Gaussian splats, not streaming video.
Versus whom: in the open world, Ant's LingBot-World (Jan 2026) was the first open model to rival Google's closed Genie 3; Astronex-World 1.0 (Sep 2026) runs 832×480 in real time on a single 48GB L20, with a lower hardware bar. Hy-Nomas differentiates on "open license + camera-control interface + a self-testable benchmark," but on the "is there evidence" column it sits exactly on the empty side — weights and test set are handed to you, just not the conclusion.
Sources
- Tencent Hunyuan Hugging Face model card and GitHub repo (2026-10-23)
- Community technical analysis on architecture, quantization route, and the benchmark gap (per official model card)