Tencent’s Hy-Nomas world model, now in Diffusers, runs on one GPU

Tencent’s Hunyuan team has published Hy-Nomas-Diffusers, a BF16 conversion of its Hy-Nomas interactive world model, so it can be loaded through the Hugging Face Diffusers library. The model, about 9B parameters, generates video one segment at a time, steered by text and camera paths. It is licensed Apache-2.0.

A 9B model that builds a world segment by segment

According to Tencent, Hy-Nomas lets users add objects, change styles and scenes, describe actions or move the camera while generation continues. Each new segment is conditioned on text, camera motion and cached history. A persistent key/value (KV) cache carries context from one segment to the next, and attention is causal across segments, so a segment does not need future ones.

The architecture combines a segment-autoregressive diffusion transformer, a text encoder taken from Qwen/Qwen2.5-VL-7B-Instruct, camera-conditioning modules and a fine-tuned PixelUMM video decoder. The transformer has 54 double-stream blocks, a hidden size of 2,048 and 16 attention heads. Camera poses enter as normalized 4 x 4 extrinsic matrices, grouped over four video frames, and are fed in through cross-attention.

The Diffusers package is the distilled version. Tencent’s example uses 4 denoising steps per segment at 704 x 1280 and 24 FPS. Its sample script generates two five-second segments, 240 frames in total, from structured JSON prompts that describe style, subjects, a segment description and timed actions. The pipeline class is HyNomasModularPipeline, and the page says installation currently goes through a branch of the Diffusers repository (diffusers[hy-nomas] from jiaxiangc/diffusers), not a released Diffusers version.

What you get, and what stays in the main repository

The folder holds the transformer, decoder and text encoder as BF16 safetensors, plus the tokenizer and a flow-match Euler scheduler. Tencent says the original Hy-Nomas repository keeps the native checkpoints, quantized variants, adapters and the Nomas-PE assets. Nomas-PE is the prompt-enhancement service behind the web demo, and this pipeline does not run it, so users must write structured prompts themselves or pass plain text.

Tencent also released Hy-Nomas-Bench, four benchmarks with 280 test cases. They cover multi-turn text interaction, camera navigation, mixed text and camera control, and spatial memory. The Diffusers page gives no scores from them.

Hardware is where the two releases differ. The Diffusers example loads everything onto one CUDA GPU, and Tencent warns that memory use rises with resolution, segment length and retained history. The main repository describes a standard setup with 8-way sequence parallelism, plus a separate single-GPU RTX 4090 configuration. That one uses INT8 transformer weights and an INT8 KV cache, an FP8 decoder, CPU offload and a 18-frame GPU history budget. Tencent’s main page states that the main configuration uses 8 denoising steps at 720 x 1280, not the 4 steps shown here.

How it stacks up against Google’s Genie 3

The obvious rival is Google DeepMind’s Genie 3, which was shown in August 2025. Secondary write-ups, such as OpenCV’s summary, describe it as generating interactive environments from text at 720p and 24 fps. Those write-ups disagree on how long a scene stays consistent: one says several minutes, against roughly 10 to 20 seconds for Genie 2 according to Design Compass, while Neurond puts it at about a minute. The numbers should be checked against DeepMind’s own material.

On paper the output format is similar: 24 FPS and roughly 720p. The difference is access. Hy-Nomas ships weights under Apache-2.0, with a standard Diffusers-style loading path, so researchers and developers can run, modify and fine-tune it. Genie 3 was presented as a model showcase, and the sources above describe no public weights.

Tencent does not compare Hy-Nomas with Genie 3 or any other world model. Neither the Diffusers page nor the main repository reports benchmark scores, generation speed or how long a world stays coherent. The main repository lists the technical report as “Coming Soon”, so quality claims cannot yet be tested against rivals. Hy-Nomas-Bench may help once outside groups publish results, but so far only Tencent’s own benchmark exists.

The release fits a recent run of Tencent open-weight uploads, among them the WeVisOne-VL vision model that we covered earlier, which also carries an Apache license.

What is still unproven

Everything about quality in the announcement is Tencent’s description, not independent testing. The page does not say how many segments can be chained before the world drifts, what a 10-second clip costs in GPU time, or how much VRAM the BF16 example needs. Exact speed depends on hardware and settings, Tencent says. Anyone who wants to try it now has two routes: the web demo linked from the model pages, or the BF16 files with the Diffusers branch on a CUDA GPU. Optional FlashAttention and MagiAttention kernels, which match the native implementation more closely, have to be installed separately.