How Much To Run AI
Open weights

Hy-Nomas: RTX 4090 setup and world-model costs

Tencent Hy-Nomas is a downloadable interactive video world model. Text instructions and camera controls steer a world across successive video segments. The official low-memory recipe targets one RTX 4090 using quantized components and CPU offload; the standard configuration uses eight GPUs. Its 24 FPS setting describes output playback, so budget GPU time from complete runs.

Content reviewed:

Hy-Nomas single-GPU workflow: Qwen text encoder, INT8 DiT and KV storage, FP8 PixelUMM decoder, CPU offload and a separate Nomas-PE service for the web demo.
Hy-Nomas runtime components, CPU offload and the separate prompt-enhancement service.

Use this model

Official model repository

Local tools and guides

Use cases

  • Build successive video segments with changing instructions and camera trajectories.
  • Evaluate the supplied single-GPU recipe before reserving a larger inference group.
  • Budget generation, decoding and the prompt-enhancement service as separate stages.

Choose offline generation or the interactive demo

Offline inference accepts a structured JSON prompt with segments, descriptions and action timestamps. It saves generated video, latents and prompt text to the chosen output directory. Use this path to check the model assets and memory setup before adding a web interface.

The full interactive web demo uses Nomas-PE, a separate prompt-enhancement service that turns user instructions into structured prompts. Include that service and its model assets in a demo budget.

RouteWhat to provision
Offline generationDiT + matching PixelUMM decoder + Qwen text encoder
Low-memory offlineINT8 DiT/KV + FP8 decoder + CPU offload
Interactive web demoGeneration stack + separate Nomas-PE service

What changes for one RTX 4090?

Tencent’s single_gpu_low_mem_inference.yaml stores DiT weights and cached keys and values in INT8, while DiT linear and attention computation remains BF16. It uses an FP8 PixelUMM decoder, limits GPU history to selected latent frames and archives earlier history on CPU. Components move to the GPU for their stage and back to CPU afterward.

Use the compatible INT8 DiT and FP8 decoder pair. Host RAM, model storage and CPU-to-GPU transfers are part of this configuration, so a 24GB GPU is only one part of the machine budget. The INT8 loader does not support a separate LoRA adapter at inference time.

The standard recipe uses eight-way sequence parallelism and distributed decoding.

Run the exact low-memory configuration

Prepare Linux, a compatible NVIDIA driver, Docker and NVIDIA Container Toolkit. Tencent recommends its prepackaged CUDA 13.0 image; the manual route uses Python 3.12+, PyTorch 2.10.0 and compiled attention kernels. Clone the official code repository and use its environment instructions before the command below.

Download the INT8 distilled DiT and FP8 PixelUMM weights from tencent/Hy-Nomas, preserving their directory layout. The Qwen/Qwen2.5-VL-7B-Instruct text encoder is another download, reused from the local Hugging Face cache. Set Hy_Nomas_Model_Path to the model repository root and expose one GPU to the launcher.

The bundled example prompts provide an initial structured workload. After the run, inspect the saved video under ./outputs/low-mem and record elapsed time and peak GPU and host memory. Use the same prompt and configuration when comparing machines.

From the code repository, after installing the official environment and assets
export Hy_Nomas_Model_Path="/path/to/Hy-Nomas-models"
bash scripts/inference/inference_from_config.sh \
  configs/experiment/single_gpu_low_mem_inference.yaml \
  dit_weight="$Hy_Nomas_Model_Path/DiT/distilled_int8/diffusion_pytorch_model_ema_0_995_int8.safetensors" \
  model.vae.checkpoint="$Hy_Nomas_Model_Path/VAE/PixelUMM_FT_FP8/diffusion_pytorch_model_fp8.safetensors" \
  inference.prompt="$Hy_Nomas_Model_Path/examples/prompts.json" \
  inference.save_path=./outputs/low-mem

Estimate cost from elapsed GPU time

The published defaults are 720 × 1280 output, 24 FPS playback and eight DiT denoising steps. Measure prompt encoding, denoising, offload transfers and decoding together, with setup and idle periods included when the rental remains billable.

For a rented setup, total cost is the sum of each machine’s hourly rate multiplied by its reserved hours, plus storage and transfer charges. Divide by the duration of accepted output only after including failed runs and retries. If Nomas-PE runs on another machine, add its reserved time too.

Keep resolution, segment length, history policy, checkpoint pair and hardware fixed when comparing runs. Use the site’s video calculator for supported hosted video models; for Hy-Nomas, use a budget based on the measured workload and reserved machine time.

ComfyUI and serving integrations

Tencent now links community integrations for ComfyUI, SGLang Diffusion and Reactor Runtime. These provide different interfaces and execution paths around the same world-model family. Follow each integration’s asset and launch instructions.

PedroMarinhoDev’s ComfyUI port uses Tencent’s INT8 weights with its own optimized kernels and an FP8 decoder. Its maintainer reports 85 seconds for a warm 10-second, 1280 × 704 fox workload on an RTX 4090, including sampling, cache passes and decoding. The separate 12GB and 16GB results were simulated by reserving memory on that 4090.

That port lists at least 32GB of host RAM and describes history growing with successive segments. Its image-start node seeds the model history using a still clip.

Check licenses across the generation stack

Tencent publishes Hy-Nomas under Apache 2.0. Its repository includes third-party notices, and the separately downloaded text encoder is another component of the deployment. Keep those licenses with redistributed assets and budget the full stack you actually operate.

Before you use it

  • The interactive demo adds prompt enhancement and serving resources beyond the low-memory offline recipe.

How to get started

Start with the official offline workload and matching quantized assets, measure the complete run, then add interactive services if the hardware budget fits.

Sources

Sources (6)

Calculators