Genie 4 vs HY-Nomas: The Better World Is Closed, the Downloadable One Is Not

For roughly 196 gigabytes you can own a world model outright. That is the size of the full HY-Nomas checkpoint set, which Tencent's Hunyuan team published under Apache 2.0 on October 23, 2026, with weights on Hugging Face and code on GitHub. Four days later, Google DeepMind announced Genie 4 — a better world model by most qualitative measures, with native audio and a two-minute coherence window, available to nobody except trusted testers and a waitlist. Between those two facts sits the most practical decision in generative world modelling right now: take the stronger model you cannot have, or the weaker one you can run tonight.

The comparison is not really Genie 4 versus HY-Nomas on quality. It is a comparison between two postures toward the same problem, and the quality gap is the price of the posture. Understanding what you are paying, and what you are paying it in, is the point of this piece.

What Apache 2.0 Actually Gets You

HY-Nomas ships as weights, not as a service. The main checkpoint is on Hugging Face as tencent/Hy-Nomas, with a Diffusers-compatible variant alongside it, and the repository carries a LoRA fine-tuning recipe plus two adapters — one for cinematic camera work, one for robotics-style footage.

That last part deserves emphasis because no closed world model offers it. If you can fine-tune on your own footage, you can move the model toward your domain: your studio's camera language, your robot's sensor characteristics, your game's art direction. Genie 4, whatever its ceiling, will produce what Google trained it to produce, and the only lever you have is a text prompt. For a team whose world model is a competitive asset rather than a tool, that difference outweighs a great deal of visual polish.

The licence matters too. Apache 2.0 is permissive enough for commercial deployment without a negotiation, which is more than can be said for Genie 4, whose commercial terms have not been published at all.

Sixty Seconds Against Two Minutes

Both models generate interactive worlds. They generate them in different ways, and the difference determines how the duration numbers should be read.

Genie 4 is a real-time world model: it renders as you steer, holds a scene for around two minutes, and generates audio — ambient sound, music, dialogue — natively in the same pass. You change the weather by typing it. DeepMind's own caveat is that details and object permanence drift over longer runs, especially after repeated complex prompt edits.

HY-Nomas generates video in sequential segments rather than in one static file, and you steer it turn by turn with text prompts and camera trajectories. A lifelong KV cache carries world consistency across those turns. Its stated ceiling is around sixty seconds of generation at 720×1280 and 24 frames per second, using eight denoising steps in the distilled configuration, with four and six also supported.

So the honest comparison is not sixty seconds versus a hundred and twenty. It is turn-based segment generation against continuous real-time rendering. Genie 4 gives you a world that responds while you are inside it. HY-Nomas gives you a world that advances one authored turn at a time, which is a slower way to work and a far more controllable one — each segment is a decision you can inspect before committing to the next.

Under the hood the two are architecturally different in a way that explains the behaviour. HY-Nomas is a diffusion transformer — fifty-four dual-stream MMDiT blocks, a Qwen2.5-VL-7B-Instruct text encoder, DCAE-V latents and a fine-tuned PixelUMM decoder, at roughly eight billion parameters by derivation from the published checkpoint rather than from an official statement. That is a large model producing video segments. Genie 4 is a real-time simulator whose internals Google has not described in any comparable detail, paired with Gemini models that ground appearance and guide the loop.

The Feature Genie 4 Has and Nobody Else Does

Audio. Genie 4 is the first DeepMind world model to generate sound natively, and as of this writing no open-weight world model in this comparison — HY-Nomas included — makes an equivalent claim. It produces ambient noise, music, and interactive dialogue in which characters respond to what is happening in the scene.

Whether that is decisive depends entirely on your use. For film pre-visualization, it is close to decisive: a shot without sound is half a shot, and rebuilding the soundscape from scratch against someone else's timing is the part of pre-vis that nobody enjoys. For robotics simulation, it is a missing input modality that limits what the environment can teach. For a game studio generating background plate footage, it is a convenience.

And for everyone, it is unmeasured. Google has published no audio benchmark, no synchronisation measurement, and its own engineers concede that audio accuracy is not solved. The claim is real and the evidence is a demo reel.

The Hardware Bill, Worked Out

This is where the open model stops being free and starts being a capital expense.

• Default deployment — HY-Nomas is designed for eight GPUs. That is a server, not a workstation.

• Single-GPU route — a documented configuration runs on one RTX 4090 using INT8 weights for the diffusion transformer, an FP8 decoder, CPU offload and a capped history buffer.

• Disk — the full checkpoint set is roughly 196 GB; the low-memory route needs about 38 GB plus the text encoder.

• Throughput — sixty seconds of output is not sixty seconds of compute. Diffusion at eight denoising steps across fifty-four transformer blocks means generation is measured in minutes per segment, and Tencent has not published latency figures.

• Genie 4 — no published hardware requirement, no published latency, no published price. Access is a subscription to Google AI Ultra via Project Genie or an early-access partner agreement.

Run the arithmetic honestly and the 4090 path is a research machine, not a production one: it will produce a segment, slowly, with degraded memory for long sequences, which is exactly the trade the capped history buffer makes. The eight-GPU path produces usable throughput and costs what eight datacentre GPUs cost, before you count the power and the person who has to keep it running.

Which is why the comparison against a hosted API is not automatic. If your volume is a few hundred segments a month, running eight GPUs to serve them is a poor use of capital, and the right answer is an API — ours or anyone's — with a per-call price and no capex. If your volume is continuous, or your data cannot leave your network, the 196 GB download is a bargain that no hosted service can match. That is the actual decision, and it has nothing to do with which model is prettier.

Two Benchmarks, Zero Scores

HY-Nomas ships something called Hy-Nomas-Bench: 280 test cases with roughly 220 reference camera trajectories. It is a real evaluation harness and, as of release, it has no published results. The technical report is described as forthcoming.

Genie 4 has no equivalent. Google published no benchmark suite, no numeric comparison to Genie 3 beyond the doubled coherence window, and no resolution or frame rate at all.

So both sides are asking to be judged on demo footage. In that situation the tiebreaker should be reproducibility, and it is not close: you can download HY-Nomas, run Hy-Nomas-Bench yourself, and publish a number. You cannot do that with Genie 4, and if you are the kind of team that needs a number before committing, the closed model is not merely inconvenient — it is disqualifying.

What We Would Actually Recommend

If you need the most convincing world available and your work is exploratory — pre-vis, concepting, prototype-level simulation — Genie 4 is the stronger capability and you should join the waitlist and design around a two-minute shot. The audio alone puts it in a different class from everything with open weights, and the Gemini-in-the-loop architecture means it can be driven by an agent rather than by hand.

If you need a world model you can own, fine-tune, audit, embed in a product, or run without sending data to a vendor, HY-Nomas is the only option in this comparison, and it is a genuinely capable one: Apache 2.0, a single-4090 path, camera control, robotics adapters, and a published evaluation suite. You are trading photorealism and sound for control and reproducibility, and for a large class of engineering work that is the correct trade.

The surrounding stack is where the two converge. Whichever model you choose, it will sit inside an agent loop — something has to write the prompts, choose the camera trajectory, check the segment and decide whether to regenerate. That loop is made of ordinary language and vision model calls, and it is where a router earns its keep: one API across 200-plus models, automatic failover so a stalled provider is a retry rather than a lost render, and provider list pricing passed through at 0% markup so a vendor price cut reaches you the day it lands. The world model is the expensive, slow-changing part of your stack. Everything around it should be the opposite.

Six months ago the interesting question was whether open world models could be competitive at all. HY-Nomas answers that they can be competitive on control while trailing on fidelity, and Genie 4 answers that the fidelity ceiling is still moving. Neither answer settles the argument, which is why the download exists and the waitlist is long.