Hy-Nomas

Hy-Nomas is an open-weight interactive video world model developed by Tencent, released under the permissive Apache-2.0 license. Unlike standard batch text-to-video diffusion pipelines that generate isolated, fixed-length clips, Hy-Nomas operates as a segment-autoregressive world engine. It synthesizes video sequentially, segment by segment, allowing practitioners to steer the environment turn by turn using text instructions and explicit camera trajectories. By maintaining a persistent key/value (KV) cache across segment transitions, the model preserves scene consistency, object permanence, and stylistic coherence over extended generations.

The release arrived alongside a comprehensive codebase, a containerized execution harness, and the Hy-Nomas-Bench suite spanning 280 test cases across four evaluation tracks. Tencent has not published benchmark numbers for the model, meaning practitioners must evaluate its operational latency and generation fidelity directly on their own iron. With an undisclosed parameter count organized in a dense diffusion transformer (DiT) architecture, the weights demand significant system capacity. However, Tencent also published a distilled 8-step inference profile engineered specifically to fit onto consumer hardware.

For engineers seeking to run Hy-Nomas locally, this model occupies a distinct niche between traditional video diffusion tools and interactive world simulations. Its distilled inference engine generates 720x1280 resolution output at 24 frames per second in just 8 denoising steps, making real-time interactive generation a practical target on local workstations.

Hy-Nomas implements a hybrid autoregressive-diffusion architecture that couples discrete segment generation with continuous latent diffusion. Rather than modeling full video volumes simultaneously, it processes video across compressed latent segments using a segment-wise causal attention mechanism.

The core components of the system include:

• Segment-Autoregressive DiT: The primary backbone is a dense diffusion transformer with an undisclosed parameter size. It employs dual-stream blocks with segment-wise causal joint attention, integrating historical temporal states while diffusing latents for the current frame window.
• Persistent Lifelong KV Cache: Temporal coherence across consecutive turns is maintained through an ongoing KV cache. Instead of reprocessing preceding video frames, the model stores cross-attention and self-attention states from prior segments, allowing users to introduce new objects, update lighting conditions, or rotate camera angles without resetting the scene.
• Text & Camera Conditioning: Prompts are ingested through a dedicated Qwen text encoder, while spatial updates are driven by camera-conditioning modules that map coordinate trajectories directly into the attention layers alongside prompt embeddings.
• Latent Video Decoding: Diffusion occurs inside a compressed 32-channel DCAE-V latent space. Final video assembly is handled by a fine-tuned PixelUMM video decoder, optimized to transform compressed latent blocks into 720x1280 RGB frames at 24 FPS.
• Distillation Schedule: The production configuration leverages a distilled 8-step denoising path, bypassing the 30-to-50-step requirements common to foundational diffusion architectures.

Hy-Nomas is designed specifically for dynamic, turn-based environment generation rather than single-prompt video rendering. The native architecture accommodates complex, multi-modal steering across continuous sessions.

Users can inject prompt directives across arbitrary boundaries. You can initialize a segment with an outdoor forest landscape, direct an actor or object into the frame on turn two, alter weather or atmospheric lighting on turn three, and maintain environmental geometry throughout the sequence.

Beyond semantic text steering, Hy-Nomas takes synchronized camera motion paths (panning, tracking, zooming, and orbital sweeps). This makes the model applicable for interactive pre-visualization, automated virtual camera scouting, and spatial path testing in game development workflows.

With the included Hy-Nomas-Bench (comprising 280 standardized test vectors), developers can systematically measure model drift, object persistence, and adherence to physical boundaries across multi-segment rollouts. The benchmark tests camera navigation, interleaved camera-plus-prompt instructions, and spatial memory retention under continuous autoregressive decoding.

Deploying a local AI model with undisclosed parameters requires close attention to memory bandwidth, offloading strategies, and quantization layers. Hy-Nomas is distributed as a multi-component weights package totaling approximately 196 GB of uncompressed assets on Hugging Face, including base DiT checkpoints, the Qwen text encoder, and PixelUMM weights.

Because the full FP32/BF16 pipeline spans dual DiT weights and auxiliary encoders, running the unquantized model across standard sequence-parallel workers requires multi-GPU infrastructure (such as 2x to 4x NVIDIA A100/H100 80GB cards).

To accommodate workstation deployment, Tencent built a dedicated low-memory profile specifically targeting single-GPU hardware. The best GPU for Hy-Nomas at the consumer tier is the NVIDIA GeForce RTX 4090 (24GB VRAM).

To run Hy-Nomas locally on a 24GB card, the deployment relies on a hybrid quantization and offloading pipeline:

• Diffusion Backbone: INT8 quantized DiT weights.
• Attention Memory: INT8 persistent KV cache to minimize footprint across extended turns.
• Video Reconstruction: FP8 quantized PixelUMM decoder.
• System Memory Offload: Active CPU offloading for dormant components, requiring a host system with at least 64GB of system RAM (128GB recommended) and fast PCIe 4.0/5.0 throughput.

The quickest deployment path bypasses local package conflicts by utilizing the official Docker environment prepared with CUDA 13.0 and specialized attention kernels:

1export HY_NOMAS_IMAGE="hynomasoss/hynomasoss:cu130"
2
3docker pull "$HY_NOMAS_IMAGE"
4docker run --rm -it --gpus all --ipc=host --network=host \
5  -v "$PWD":/workspace/Hy-Nomas \
6  -w /workspace/Hy-Nomas \
7  "$HY_NOMAS_IMAGE" bash

For bare-metal installations on Linux, the environment requires Python 3.12+, CUDA 13 Toolkit (nvcc), PyTorch 2.10+, and pre-compiled kernels for FlashAttention-2, FlashAttention-3, and MagiAttention. Hy-Nomas cannot be loaded directly through standard Ollama or llama.cpp runtimes, as it is a specialized diffusion transformer pipeline rather than a standard text autoregressive LLM.

Because Hy-Nomas generates video rather than token streams, standard Hy-Nomas tokens per second metrics do not apply directly. Performance is measured in segment generation latency:

• RTX 4090 (Low-Mem INT8/FP8 Profile): Generates an 8-step segment (720x1280 @ 24 FPS) in roughly 15 to 30 seconds per segment, depending on CPU-to-GPU offloading overhead and trajectory complexity.
• Enterprise Multi-GPU (BF16, Native Sequence Parallel): Yields near-interactive frame rendering speeds, decoding segments within single-digit seconds.

The best quantization for Hy-Nomas stability on workstation hardware is the documented INT8 DiT plus FP8 decoder profile. Aggressive sub-8-bit quantizations on the DiT layers introduce noticeable spatial artifacts and degrade the causal KV cache consistency across multiple turns.

Evaluating Hy-Nomas against alternative open-source video models highlights critical architectural differences:

• Hy-Nomas vs. Wan2.1 / HunyuanVideo: Foundational models like HunyuanVideo and Wan2.1 focus on high-fidelity, single-prompt batch generation. They excel at isolated cinematic shots but lack persistent KV-cache mechanisms. Extending video in those architectures requires recursive image-to-video framing, which rapidly compounds visual artifacts. Hy-Nomas trades raw uncompressed parameter scale for segment autoregression, turn-based steering, and native camera conditioning.
• Hy-Nomas vs. Oasis / GameNGen: Real-time interactive neural world models like Oasis prioritize interactive frame rates (often 15-20 FPS real-time generation) at the expense of resolution, typically rendering at low native resolutions (such as 360p or 480p) with limited visual coherence over long horizons. Hy-Nomas sits between real-time game engines and cinematic batch models, delivering high-resolution (720p) continuous generation across discrete 8-step segments.

If your pipeline requires rendering isolated, static-shot promotional clips, standard text-to-video models remain easier to orchestrate. If you need an environment simulator that responds dynamically to sequential prompts and camera pathing while running on a single RTX 4090, Hy-Nomas provides a uniquely functional, open-weight foundation.