Apollo World vs Hy-Nomas: Pixels That Must Look Right, Sensors That Must Be Right

Show two people the same ten-second clip of a car approaching a junction and both will agree it looks fine. Show it to a planning stack and only one version is usable, because the other has a pedestrian whose shadow points the wrong way, a kerb that is 40 centimetres too high, and no lidar return at all. That gap is the whole difference between Apollo World and Hy-Nomas. Hy-Nomas is Tencent Hunyuan's open-weight interactive video world model: an 8-billion-parameter model that generates 720p video from a text prompt and a camera path, released under Apache 2.0 in late October 2026. Apollo World is Baidu's agent-driven simulator, disclosed on 3 November 2026, which emits camera and lidar observations of a virtual city in which the traffic reacts to the vehicle under test. One is judged by the eye. The other is judged by whether the car gets better at driving.

Both are legitimately called world models. Only one of them can be downloaded this afternoon, and only the other one is designed to feed a perception stack. That trade is the article.

The fidelity trap

Video world models optimise for what a viewer will accept. That is not a criticism; it is the correct objective for a generation model, and it is why Hy-Nomas can produce a convincing walk through a street it has never seen. But perceptual plausibility and physical consistency are different targets, and they come apart precisely where autonomous driving lives.

A model trained to predict the next frame learns that cars slow near junctions and that shadows fall away from a light source. It does not learn, from pixels alone, that the distance between two lane markings is 3.5 metres, that a lidar pulse returns from a wet road at a particular intensity, or that a pedestrian who has already begun to step out cannot teleport back to the kerb when the ego vehicle brakes. Ask it for a sensor stream and it will produce something that looks like a sensor stream. Whether the numbers in it are metric is a separate question that the training objective never asked.

Baidu's launch material is built around that objection. Wang Liang, chief architect of Baidu's Intelligent Driving Group, said on stage that a world model should not be a painter producing impressive video but an examiner of driving safety, and the "2+3" rubric that followed is explicitly a filter for models that look right without being right. Causality and interactivity are the gates; physical accuracy, persistence and real-time performance are the graded properties. Whatever else the rubric is, it is a description of the failure mode of a video generator applied to a driving stack.

What Hy-Nomas actually ships

Hy-Nomas is a segment-autoregressive diffusion transformer, not a single-shot video generator. It produces variable-length segments conditioned on a cached history, which is how it maintains coherence as the camera moves: the "world memory" is a KV cache carried forward, refreshed as new segments are generated.

• Size — roughly 8 billion parameters across 54 dual-stream MMDiT blocks with a hidden size of 2,048.

• Inputs — free-form text plus a continuous camera trajectory, with an optional prompt enhancer, Nomas-PE, that rewrites loose prose into structured JSON captions.

• Output — 720 by 1280 at 24 frames per second, using a fine-tuned PixelUMM decoder over DCAE-V latents, in eight distilled denoising steps.

• Weights — tencent/Hy-Nomas on Hugging Face under Apache 2.0, around 196GB, with code on GitHub and a hosted demo.

• Hardware — a default deployment wants eight GPUs. A single RTX 4090 path exists using an INT8 DiT and KV cache, an FP8 decoder and CPU offload, at roughly 38GB.

The technical report, arXiv 2610.36729, appeared on 27 October 2026 and describes the trick that makes multi-turn interaction work: structured captions are aligned with an action-aware causal attention pattern, so descriptive context persists across segments while action instructions are confined to the window in which they were issued. The evaluation is a 280-case benchmark across four tracks — multi-turn text, camera navigation, interleaved text and camera, and spatial memory — with human pairwise comparisons against Genie 3, LingBot-World v2 and Happy Oyster. Hy-Nomas wins on interaction and prompt following, with a reported win rate around 65% on prompt following, and loses to Genie 3 on long-term spatial memory. Those are the authors' comparisons on the authors' benchmark, and they are the most checkable claims in this article because the weights are public.

What Apollo World does not publish

Apollo World's disclosure was a capability briefing, and the absences are as informative as the claims. There is no parameter count, no architecture, no paper, no benchmark table with named baselines, no weights, no API and no price. What Baidu did say is that the model is already supporting Apollo Go's global testing and validation, that it can cover around 100,000 kilometres of simulated driving per day — framed as roughly three years of continuous testing by a safety-driver-equipped car — and that Apollo Go operates in 28 cities with more than 23 million rides, 350 million autonomous kilometres and more than 240 million fully driverless kilometres.

The 100,000-kilometre figure is the one to hold loosely. It is a company number with no external measurement behind it, and the comparison it rests on — three years of a test car running continuously — depends on an assumed annual mileage that Baidu did not state. That does not make it wrong. It makes it unverifiable, which is a different and more useful thing to know.

The design claims, by contrast, are specific enough to argue with. Every traffic participant is an agent that decides rather than a prop that follows a script. Camera and lidar are generated in real time. Trips close end to end, from accepted order to delivered passenger, across arbitrary origins and destinations. And the whole thing is bounded by a decision cycle of one thirtieth of a second, about 33 milliseconds, which is what makes the persistence and latency properties load-bearing rather than decorative.

Applying Baidu's own rubric to Hy-Nomas

The interesting exercise is not to compare the two on marketing terms but to take Apollo World's five properties and point them at the model that actually exists.

Causality: Hy-Nomas passes, in the sense Baidu means. It is autoregressive over segments and conditions on history, so the future is generated from the present rather than assembled and back-filled. Interactivity: partially. It responds to text instructions and camera moves, but those are director's commands, not the actions of other agents. Nothing in the scene reacts to anything else in the scene. Physical accuracy: unmeasured, and unmeasurable from the published work, which evaluates against human preference rather than metric ground truth. Persistence: this is the documented weak spot — the paper itself reports trailing Genie 3 on long-term spatial memory, which is exactly the drift property Baidu calls out. Real-time: 24 frames per second in eight distilled steps is fast for a diffusion model, but the reference deployment is eight GPUs, and the 33-millisecond budget is not in the same conversation.

None of that makes Hy-Nomas a bad model. It makes it a different instrument. Judged as an interactive video world — the thing it was built to be — it is competitive with the best available and it is open, which almost none of its competitors are. Judged as a driving simulator, it fails four of Baidu's five tests, and would fail them even if the weights were twice as good.

What it costs to run each

Hy-Nomas has a price you can calculate. The weights are free under Apache 2.0; the cost is the machine. A single 4090 running the quantised path needs about 38GB of memory once CPU offload is counted, and produces 720p at 24 frames per second. The eight-GPU reference deployment is the version the benchmarks describe. If you are generating a few hundred clips for a research question, one card is enough; if you are generating a training corpus, the electricity bill is the licence fee.

Apollo World has no calculable price because it is not for sale. The only honest statement is that Baidu has published no commercial terms, and the deployment described is inside its own robotaxi operation.

For teams building around either, the routing question is separate from the model question. Neither Apollo World nor Hy-Nomas is in OrcaRouter's catalogue — one is internal to Baidu, the other is a video model you host yourself — and this piece is not going to pretend otherwise. What we do carry is the layer that evaluates and orchestrates this kind of output: a vision-language judge that scores generated clips, a planner that turns a research question into a batch of prompts, a fallback model when the primary is rate-limited. That is one API over more than 200 models at 0% markup, with the provider's list price passed through, which matters most when you are running thousands of judge calls against a 196GB model's output.

Which one belongs in your stack

If you want to explore a generated environment, direct a camera through it, and iterate on a scene in a browser, Hy-Nomas is the better answer and it is available now. The open licence is not a formality: it lets you fine-tune, quantise and ship, which is more than can be said for most interactive world models, including the ones it beats on prompt following.

If you are validating a driving policy, Hy-Nomas is not a substitute for a simulator and will not become one through better prompting. The right use of Apollo World is as a specification: the 2+3 rubric is a set of questions you can put to any simulator vendor, and the persistence question in particular — what is the drift over a ten-hour run — is the one that separates a test rig from a demo reel.

The convergence to watch is a video world model that emits metric depth and lidar alongside RGB, trained with a physical-consistency loss rather than a perceptual one. Hy-Nomas's authors are already conditioning on camera trajectories in a geometrically meaningful way; adding a sensor head is the obvious next step and nobody has published it. Until someone does, the division is clean: Hy-Nomas makes worlds you can look at, Apollo World makes worlds a car can be graded in, and the gap between those two sentences is where the next year of work sits.