Tencent Hy-Nomas: An Open Interactive Video World Model You Can Steer by Text and Camera

Tencent's Hunyuan team has open-sourced Hy-Nomas, a video world model you steer as it plays out. According to the Hugging Face model card, it "generates a world one video segment at a time, allowing users to introduce objects, change styles and scenes, describe actions, or control the camera as generation continues." Weights, code and a benchmark are all Apache 2.0, and the model card lists a single RTX 4090 configuration. It was published on October 23, 2026.

This is a developer release, not a polished product. The technical report is marked "coming soon," no benchmark scores are published, and the standard setup wants eight GPUs. But the combination of an open license, a camera-control interface and a purpose-built benchmark makes it one of the more usable open world models so far.

TL;DR: the questions people ask first

How does Hy-Nomas work?

The model card describes a stack with four main parts:

• A segment-autoregressive diffusion transformer. Diffusion runs inside a compressed video latent space, and autoregression runs across segments. Each new segment is denoised from noise while attending to a persistent key-value (KV) cache of earlier segments. Attention is causal across segments, so future ones are not needed.
• Text conditioning. The language component of Qwen2.5-VL-7B-Instruct acts as the text encoder, so you can change the prompt every segment, for example "glowing blue flowers open beside the path."
• Camera conditioning. Camera extrinsic matrices are fed in through cross-attention in every transformer block, so a trajectory file or keyboard-style controls (W, A, S, D, arrow keys) steer the viewpoint.
• A PixelUMM video decoder. Latents use a DCAE-V representation with 32x spatial and 4x temporal compression, and a fine-tuned PixelUMM decoder turns them into pixels.

The transformer itself has 54 dual-stream MMDiT blocks, hidden size 2,048, 16 attention heads, and 3D rotary position encoding, per the card's architecture table. For a reader new to the idea, a "world model" here means a generator that keeps a scene consistent while you act inside it, much like a game engine, except the frames are generated rather than rendered. Our coverage of Skywork's Matrix-Game 4.0 and Odyssey-3 covers sibling systems.

What can you actually do with it?

According to the card, you can:

1. Start a world from a structured prompt. Offline inference reads a JSON file with a style, a perspective (such as third-person), a subject list and timed segments, each with a description and actions.
2. Change the world mid-stream. Later segments carry a description_change field that evolves the scene, such as adding objects or switching style.
3. Drive the camera. Pass a camera trajectory file (camera-to-world matrices and intrinsics) alongside the text.
4. Add LoRAs. Optional adapters include a cinematic look and a robotics adapter trained on the RealSource-World robotic-arm dataset to improve simulated manipulation scenes. The INT8 loader does not support loading a separate LoRA at inference time.
5. Use the web demo. The online demo lets you start a world, continue it with text or camera paths and download the result. It requires a separate prompt-enhancement service, Nomas-PE, that is not needed for offline use.

The robotics adapter is the part to watch for physical-AI readers. World models are increasingly used to generate training scenes for robots; see NVIDIA's Cosmos 3 surgical simulator and Naver's LaLa world model for the same direction.

A filmstrip with one frame turning green, standing for segment-by-segment generation in the Hy-Nomas video world model

Hy-Nomas-Bench: what the 280 test cases measure

Alongside the model, Tencent released Hy-Nomas-Bench, described as "280 test cases for interaction controllability in video world models." It has four suites:

Details worth knowing:

• Camera navigation actions last 2 to 4 seconds and are paired with an equal-length inverse movement, so the generated video should end where it started.
• The spatial-memory suite has first-person scenarios that go out and back four times around a distinctive static object; you compare frames at matching camera poses to see whether the object stayed consistent.
• Scenarios cover real-world, game, CG and animated environments; prompts are mostly English with a few Chinese or mixed-language ones.
• The release contains 220 reference camera trajectories as NPZ files, plus prompts and interaction sequences. It does not include videos, extracted frames or scores, and Tencent says scoring is "a separate step."

That last point matters. The benchmark gives the community a shared test set, but there is no published leaderboard or scoring code in the release, so numbers from different groups will only be comparable if they agree on a scorer. Treat any "Hy-Nomas scores X" claim that appears soon as unverified until Tencent's report lands.

How to run it

The model card recommends Tencent's prebuilt Docker image (hynomasoss/hynomasoss:cu130) with a Linux host, NVIDIA drivers and the NVIDIA Container Toolkit. Manual installation uses uv, Python 3.12 or newer, the CUDA 13.0 toolkit and PyTorch 2.10, and compiles several attention kernels.

Two launch modes are documented in the GitHub repository:

Honest caveats from the card itself:

• The standard launch "requires eight GPUs in one inference group." The RTX 4090 path trades speed for memory: CPU-GPU transfers add latency, and system RAM must hold the offloaded models and a growing history archive.
• The card says generation speed depends on hardware, resolution, segment length and history, and does not give a frames-per-second figure for generation; the 24 FPS is the playback rate of the output.
• The history budget caps context, not the memory used by the current segment.
• The prompt format is verbose. You write timed segments with subject lists, not a single sentence, unless you use the web demo's prompt enhancer.

A sensible first test

If you want to judge Hy-Nomas yourself, start small and measure the things the benchmark is built around:

1. Try the web demo first to see how text changes and camera moves behave before committing GPU time.
2. Run one Bench-4 spatial-memory case with the offline pipeline and compare the first and last frames at the same camera pose. If a distinctive object changes shape or disappears, that is the failure the benchmark exists to expose.
3. Run one Bench-2 camera case and check that the inverse movement returns to the starting view.
4. Write down your settings. The benchmark card asks users to report case IDs, prompt representation, generation settings and any exclusions, so your results can be compared with others.
5. Watch memory growth on long runs. Full-history mode keeps adding cache, while the single-GPU mode bounds GPU history but grows a CPU archive.

None of these steps needs the unpublished technical report, and together they give a more honest picture than a curated demo clip.

How it compares with other open world models

We have not run Hy-Nomas ourselves, and Tencent has not published head-to-head numbers, so this table compares stated design goals, not quality.

Why it matters

Three reasons stand out. First, an open license: Apache 2.0 for weights, code and benchmark lets teams fine-tune and ship. Second, the interface: multi-turn text plus camera trajectories is closer to how game and robotics developers want to drive a simulator than single-shot text-to-video. Third, the benchmark: spatial memory and interleaved control are exactly where interactive generators tend to break, and having 280 shared cases makes comparisons possible.

It also continues a run of Tencent open releases, including the Hy-Memory-30B-A3B memory writer and Youtu-Parsing-Omni.

What is not confirmed yet

• Tencent's technical report is listed as coming soon, so training data, compute and ablations are unknown.
• No quantitative results versus other world models are published.
• The robotics LoRA's claims about improved simulation have no published measurements.
• Real-world 4090 speed and the length of coherent worlds you can generate are untested by third parties.
• The model card's acknowledgements mention several recent open models, but the card does not say how any of them were used.

Related reading

• Skywork Matrix-Game 4.0: open world model, playable games on RTX 5090
• Odyssey-3 and Physics-IQ world model results
• Naver LaLa world model with language actions
• NVIDIA Cosmos 3 generative surgical simulator
• Tencent Hy-Memory-30B-A3B
• Tencent Youtu-Parsing-Omni

Details are accurate as of October 23, 2026, from the Hugging Face model and dataset cards; Tencent may update the repositories.