PAN 2 vs Hy-Nomas: Same Licence, Same Architecture Family, Two Definitions of "Action"

This is the only pair in the current crop of world models where the comparison is a real choice rather than a category error. PAN 2 is the Institute of Foundation Models' 29-billion-parameter world model, Apache-2.0 on Hugging Face since 7 October 2026, code on 5 November, Day-0 vLLM-Omni support on 10 November, which turns a starting frame and a described action into up to twenty seconds of 704×1248 video at 24 frames per second. Hy-Nomas is T​encent Hunyuan's interactive video world model, Apache 2.0, roughly 8 billion parameters, described in arXiv 2610.36729 on 27 October 2026, which turns a text prompt plus a continuous camera trajectory into 720p video at 24 frames per second in eight distilled steps. Both are open-weight. Both are diffusion transformers generating video from a conditioning signal. Both are downloadable this afternoon. Anyone choosing between them cannot use licence, availability or category as a tiebreak, which is exactly what makes it worth doing carefully.

The tiebreak that does work is a sentence neither team puts in its abstract: what counts as an action. Hy-Nomas's control signal is where the camera goes. PAN 2's control signal is what happens in the scene, written down by a language model as a sequence of events bound to actors. Those are different enough that a prompt which works well on one is close to meaningless on the other, and if you only learn one thing from this comparison, that is the one.

Two conditioning interfaces, not two quality levels

Hy-Nomas conditions on text plus a camera path. The path is continuous and geometric: you tell it where the viewpoint moves, and the model generates a segment conditioned on a cached history so the world stays coherent as you travel. Its prompt enhancer, Nomas-PE, rewrites loose prose into structured JSON captions, which is the same instinct PAN 2's prompt-rewriter follows from the other direction — both teams concluded that free text is too vague to condition a diffusion transformer on.

Where they part is what the structured form describes. Hy-Nomas's structure is spatial and viewpoint-centric; the thing being modelled precisely is the camera's motion through a scene. PAN 2's structure is temporal and actor-centric: events in order, each action bound to the actor performing it, plus camera motion as one field among several. Read IFM's own framing and the emphasis is unmistakable — the caption describes what happens after an action, and the demonstration that made the model famous is a strawberry dropped into a bowl where swapping the fruit or the liquid changes the splash.

• Conditioning — Hy-Nomas takes free text plus a continuous camera trajectory; PAN 2 takes an optional first frame plus a structured, time-grounded action caption.

• What "action" means — Hy-Nomas: a camera move, i.e. where you look; PAN 2: an event with an actor, i.e. what happens.

• Segment structure — Hy-Nomas is segment-autoregressive over a cached KV history, so context carries forward across segments; PAN 2 simulates each clip from at most one starting frame plus text, with no carried state between clips.

• Output — Hy-Nomas: 720×1280 at 24 fps in eight distilled steps; PAN 2: 704×1248 at 24 fps across fifty sampler steps at guidance 2.5.

• Size — Hy-Nomas: roughly 8B across 54 dual-stream MMDiT blocks at hidden size 2,048, plus a fine-tuned PixelUMM decoder over DCAE-V latents; PAN 2: a 29B MMDiT plus a frozen Qwen3.5-9B text encoder and a Wan 2.2 VAE.

• Hardware — Hy-Nomas: eight GPUs in the reference deployment, with a single RTX 4090 path using an INT8 DiT, KV cache, FP8 decoder and CPU offload at roughly 38GB; PAN 2: eight H200s in BF16 as documented default, with a recipe putting thirteen seconds of video at about 119 seconds on four B300s.

• Weight size — Hy-Nomas: around 196GB on Hugging Face as tencent/Hy-Nomas; PAN 2: 86.17GB across sixty files as IFM/PAN2.

Seven dimensions and not one of them is a score. What they describe is a design divergence: Hy-Nomas put its parameters into holding a world together as you move through it, PAN 2 put its parameters into predicting what an action does. A camera path is cheap to specify and hard to fake; an event sequence is expressive and expensive to generate.

Where the difference bites: three tasks, both models

Abstract contrasts are cheap. These are concrete.

Walk me through a flat. Hy-Nomas wins outright. You want a continuous path through a corridor and into a kitchen, with the room behind you staying the room behind you. That is segment-autoregressive generation over a cached history, which is precisely what its architecture is for, and PAN 2 cannot do it: it renders a clip, and the next clip starts from a frame you supply with no memory of what came before. Chaining PAN 2 clips to make a walkthrough produces the drift its own model card warns about, where appearance changes across clips in long interactions because each is simulated from at most one starting frame.

I open the drawer and the cup falls. PAN 2 wins, and this is not close. The request is an event with an actor and a physical consequence. Hy-Nomas's conditioning has no field for "drawer opens" separate from the camera — you can put it in the text prompt, and its prompt enhancer will structure it, but the model's precision budget was spent on viewpoint rather than on contact and consequence. PAN 2's caption contract exists precisely to bind actions to actors and order them in time, which is why IFM shipped a validator that checks descriptions against the text encoder's budget of 1,480 tokens soft and 1,536 hard rather than letting prose through.

Generate two hundred variants of a manipulation scene for a dataset. Both are wrong, differently. Hy-Nomas will give you two hundred coherent camera moves through two hundred scenes and very little control over whether the arm actually grasps anything. PAN 2 will give you two hundred action-conditioned clips and, per its own card, about two in three plans will contain at least one step that is not carried out — so a third of your dataset will be unusable unless you build a checker. Neither is a data engine out of the box; the difference is whether the failure is geometric drift or a dropped step, and which of those your downstream task can tolerate.

That third task is also the one that determines your hardware. Two hundred clips of eight-step 720p on a single quantised 4090 is an afternoon. Two hundred clips on a four-B300 node at 119 seconds per thirteen seconds of video is a couple of days, and that arithmetic is the real reason the two models are not interchangeable even where their capabilities overlap.

The same failure with two different names

Both cards admit the same weakness and neither fixes it: a long-horizon world model loses its world. PAN 2 says appearance drifts across clips and names perceptual memory as the missing piece. Hy-Nomas addresses it directly with a carried KV cache and a causal attention pattern that keeps descriptive context alive across segments while confining action instructions to the window in which they were issued — and then reports in its own evaluation that it still trails Genie 3 on long-term spatial memory.

That last admission is more interesting than a win. It says the memory mechanism works well enough to beat Genie 3 and LingBot-World v2 on interaction and prompt following — around a 65% win rate on prompt following across a 280-case benchmark covering multi-turn text, camera navigation, interleaved text and camera, and spatial memory — and still loses on the track that measures whether the world stays put. Those are the authors' comparisons against the authors' chosen baselines on the authors' benchmark, and like every number in this category they are unreproduced. What they are not is vague: a model that publishes where it loses is a model you can plan around.

The useful translation for anyone building on either: persistence is not a property you can assume and check later. It is the first thing to measure on your own scenes, with your own tolerance for drift, before you commit a pipeline to a model. Neither vendor publishes a time-to-failure curve — the duration at which a scene stops being trustworthy — and that missing number matters more than any benchmark position.

Cost is not a rounding difference, it is the decision

Set the capability argument aside for a moment and look at what it takes to run each one for a week.

Hy-Nomas's single-4090 path is the most consequential fact in this comparison and it is easy to miss under the eight-GPU reference deployment. An INT8 DiT with a KV cache, an FP8 decoder and CPU offload fits in roughly 38GB, which means a studio with one workstation-class card can generate 720p interactive video without renting anything. The eight-GPU configuration is what the benchmark describes; the 4090 configuration is what a reader can actually do tonight.

PAN 2 has no comparable path published. Its documented default is eight H200s in BF16 with VAE patch parallelism across four tiles, and the serving recipe that everyone quotes — thirteen seconds of 704×1248 in roughly 119 seconds — assumes four B300s. Worse, the sampler is expensive by design: fifty steps with guidance of 2.5 applied at every step is a hundred forward passes through a 29B transformer per clip, which is why classifier-free guidance has to be split across ranks and sequence parallelism spread across more. There is no quantised single-card story in the release, and pretending otherwise would be inventing one.

So the honest cost sentence is this: Hy-Nomas costs a GPU you probably have, PAN 2 costs a GPU node you probably do not, and no amount of quality difference in either direction changes that for a small team.

The routing angle here is narrow and worth stating precisely, because both of these are models you host yourself and neither is in the OrcaRouter catalogue. What is routable is the part both architectures added because free text was not enough: the prompt-enhancer stage. Hy-Nomas's Nomas-PE rewrites prose into structured JSON captions; PAN 2's prompt-rewriter holds a system prompt, a fixed negative prompt and a validator against the encoder's token budget. Those are language-model calls, and they are the stage where a different model measurably changes what the renderer receives. IFM ranked four reasoners for its caption stage — Claude Opus 5.5, GPT-6 Astra at medium effort, Kimi K3, Qwen3.8-Flash — and all four are in the catalogue, so the rewriting half of either pipeline can be swapped and compared through one key at provider list price with 0% markup, with failover when a rewrite call times out mid-batch. The 29B and the 8B renderers stay on your hardware. That is the whole claim.

What the two model cards actually contain

Both releases put a card in public and the cards differ in what they are willing to admit.

Hy-Nomas's report, arXiv 2610.36729 on 27 October 2026, 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 — judged by human pairwise comparison against Genie 3, LingBot-World v2 and Happy Oyster. It wins on interaction and prompt following, with a reported win rate around 65% on prompt following, and it loses to Genie 3 on long-term spatial memory. The weights are around 196GB as tencent/Hy-Nomas under Apache 2.0, with code on GitHub and a hosted demo.

PAN 2's card is longer and less flattering. The architecture is generative latent prediction from Critique of World Models (arXiv 2507.05169), and the same document that describes it lists the failures: about two in three simulated plans drop at least one step, appearance drifts across clips because there is no carried state, wording of the observation and action changes the outcome, text expresses "open a drawer" well and "place each finger here" badly, a simulation can progress while violating physics, and simulating plans that should fail is not yet measured.

Read the two side by side and the difference in disclosure style is obvious. Hy-Nomas reports a win rate and a loss; PAN 2 reports a failure rate and a list. Neither is an audit, and both were written by the teams that built the models. But a card that tells you where the model loses is a document you can plan against, and both of these qualify — which is more than most of their closed competitors offer.

Which one, and what would change the answer

Pick Hy-Nomas if your problem is exploration or content: a world someone walks through, a camera move someone directs, a coherent 720p scene from a prompt on a single card. Its memory mechanism, its camera conditioning and its 4090 path all point at the same user — someone who needs to see a place and move through it, cheaply, now.

Pick PAN 2 if your problem is a decision: an agent that needs to preview what an action does, a planner ranking candidate plans, a research question about whether a structured description of an event is sufficient to generate its consequence. Its caption contract, its published harness and its Apache-2.0 weights all point at a user who needs to be argued with honestly and has the hardware to pay for it.

Two things would break the tie in a way neither vendor currently allows. A Hy-Nomas-style memory carried across segments on a model with PAN 2's action conditioning would be a meaningfully better product than either. A published failure study — PAN 2 simulating plans that should fail, Hy-Nomas measuring drift over ten minutes rather than one — would answer the question both cards leave open. Until one of those lands, the choice is not about which model is better. It is about whether your world is something you move through or something that happens to you.