PAN 2 vs Agora-2: One Simulates Consequences, the Other Keeps a World Running

Twenty people in one world sounds like more than one person imagining what happens next. That is the whole confusion between PAN 2 and Agora-2, and it is worth untangling before anyone benchmarks them against each other. Agora-2 is Odyssey's playable research preview, published 21 September 2026, which supports twenty participants in a shared simulated environment — four of them human controllers, sixteen of them learned policies — and keeps their world consistent even when a player looks away. PAN 2 is IFM's 29-billion-parameter open-weight simulator, released as 86.17 gigabytes of Apache-2.0 checkpoint, which takes a frame and a described action and returns up to twenty seconds of video predicting the consequence. One is a server for a world. The other is a question you ask a world. The overlap in vocabulary is nearly total and the overlap in function is close to zero, which is exactly why the comparison keeps getting made badly.

The number that travels is not the number that matters

Coverage of Agora-2 leads with the twenty, and Odyssey's own post encourages that. The composition is more interesting than the count. Four concurrent human controllers supply movement and action input; sixteen autonomous entities are driven by monster and companion policies that Odyssey trained on the same corpus the world was learned from — captures of Diablo II. So the session is not twenty people in a room. It is a rendering problem with a very small state machine at its centre, dressed as a game server.

That architecture is what makes Agora-2 a different kind of object from PAN 2. Odyssey separates the thing that decides what happens from the thing that decides what it looks like, and the size gap between them is the most striking number in the report: the simulation model has 4,457,777 parameters — roughly 4.5 million, four layers wide, fourteen movement branches, a four-step history window — while the renderer is a flow-matching transformer of about one billion parameters running five solver steps per update, conditioned on a 342-token prefix describing the map, entities, transient effects, camera and roster.

PAN 2 does the opposite. It spends almost all of its capacity on perception: a 29B multi-modal diffusion transformer predicts the frames, and the part that decides what happens is not a learned simulator at all but a language model writing a structured, time-grounded caption of events, actions and camera motion. Agora-2 puts its intelligence in a tiny explicit state representation. PAN 2 puts its intelligence in the renderer and delegates the reasoning to text.

• Decides what happens — Agora-2: a 4.46M-parameter learned simulation model with explicit entity state; PAN 2: any instruction-following language model writing a world-state caption.

• Renders it — Agora-2: a ~1B flow-matching renderer at five solver steps per update; PAN 2: a 29B diffusion transformer at fifty steps with guidance on every step.

• Time horizon — Agora-2: an open-ended session that persists; PAN 2: up to 20 seconds per clip, at most one starting frame, drifting across clips.

• Input — Agora-2: movement and action commands from controllers plus learned policies; PAN 2: an optional first frame plus a structured action description in text.

• Participants — Agora-2: 4 humans and 16 policies in one shared state; PAN 2: none, it is a batch generator.

• Access — Agora-2: no weights, no repository, no licence and no price; PAN 2: Apache-2.0 weights, ungated, with GitHub code and a vLLM recipe.

Persistence is the thing Agora-2 actually solved

Strip away the multiplayer framing and Agora-2's real contribution is a sentence in Odyssey's post: because entity properties are retained independently of any particular view, they remain available when an entity is out of frame. A video generator that conditions only on recent frames loses the monster the moment you turn around. A world model that keeps explicit state does not. That is the design decision that separates Agora-2 from every video model with multiplayer ambitions, and it is why twenty participants is a state-management claim rather than a rendering one.

PAN 2 makes no such claim, and its model card says so. Each clip is simulated from at most one starting frame plus text, so appearance drifts across clips in long interactions, and the card names perceptual memory maintained alongside the text-based state as the thing that would be needed to fix it. In other words: PAN 2 knows it has Agora-2's problem and has not solved it, and Agora-2 solved a version of it by keeping the world small enough to hold in an explicit tensor.

Read together, the two designs bracket the field. If your world can be described in 4.5 million parameters — a map grid, an entity roster, a set of movement branches — persistence is tractable and you get a session. If your world is open-domain physical consequence, you need a renderer that understands materials and contact, and persistence becomes an open research problem the moment the camera moves.

What a session costs, and what a simulation costs

Neither number is published as a price, because neither model is sold. Both are published as hardware, which is more useful.

Odyssey's serving plan for a four-player session is four NVIDIA RTX PRO 4500 cards: one renderer per GPU, with one of those GPUs also hosting the simulation model and the agent policies, targeting fifteen latent updates per second against thirty displayed frames per second. The report is candid that those two figures are targets, not measurements — section 8 lists sustained frame rate and input-to-display latency during live multi-agent interaction as quantitatively unevaluated. The training bill is itemised too: sixty thousand updates on 4,232,000 segments, then sixty thousand more on 4,332,800, run on 32 B200 GPUs.

PAN 2's equivalent is the vLLM recipe: thirteen seconds of 704×1248 video in roughly 119 seconds on four B300 GPUs, with a documented default of eight H200s in BF16. Fifty sampler steps with classifier-free guidance on every step — a hundred forward passes through a 29B transformer per clip — is why guidance is parallelised across two ranks and sequence parallelism across two more.

The two configurations are of similar scale, and that is where the resemblance ends. Four cards buy a session that keeps running. Four cards, on PAN 2, buy two minutes per thirteen seconds of rendered consequence, and the cost of a plan is the cost of every candidate the planner considered. Odyssey's number is a throughput floor for an experience. IFM's is a per-question price.

Only one of these can be built on today

This is the part where the comparison stops being architectural.

Agora-2 is a research preview with a playable demo at Odyssey's own address and a technical report. There are no weights, no repository, no licence, no inference API and no price. Odyssey frames the trajectory toward multi-agent interaction inside foundational world models as future work, with PROWL as the research thread and Odyssey-3 as the eventual destination. Practically: you can play it, you cannot build on it, and no third party can serve it to you either.

PAN 2 is the opposite arrangement. Weights on Hugging Face since 7 October 2026, ungated, Apache-2.0 on its own code and weights; GitHub repository with inference and training scripts since 5 November; Day-0 vLLM-Omni support merged 10 November; a guardrail distributed as a separate repository with its own false-positive numbers. IFM's announcement was a single post on its own X account — "Introducing PAN 2: an open foundation world model for physical intelligence" — with no blog entry, paper link, hosted endpoint or pricing page. That is a thin announcement wrapped around a thick artefact, which is the inverse of Agora-2's polished post around a closed demo.

If your question is "what can I put in production this quarter", the answer is neither, but for different reasons. Agora-2 has no path from demo to deployment that Odyssey has described. PAN 2 has a path — the weights and the recipe — and a limitation that decides whether you should walk it: about two in three simulated plans contain at least one step that is not carried out, per the model card's own limitations section.

The hybrid neither lab has built

The interesting engineering question is not which model wins. It is what a system looks like that uses each for what it is good at, and neither vendor ships it.

Picture an agent that needs to plan inside a persistent world. Agora-2's design shows how to hold the world steady: explicit state, a structured conditioning prefix, policies small enough to run alongside a renderer. PAN 2's design shows how to answer "what happens if" about a physical interaction: hand the question to a language model, get a structured caption, render the consequence, look at it. A stack that kept Agora-2's state and asked PAN 2 for the fine-grained physics inside a cell of it would be strictly better than either, and would require both labs to open something they have not.

What you can do today is the half that does not need either model. PAN 2's planner is defined as any instruction-following language model, and IFM ranked four of them by full-plan completion: Claude Opus 5.5, GPT-6 Astra at medium effort, Kimi K3, Qwen3.8-Flash. All four are on OrcaRouter, so the caption stage of a PAN 2 pipeline can run through one endpoint with automatic failover and each provider's list price passed through at 0% markup, swapped per experiment rather than per contract. OrcaRouter does not host PAN 2 and cannot route it — the 29B renderer stays on your GPUs — and it certainly cannot host Agora-2, which nobody can. What a router can do is keep the reasoning half elastic while the simulation half is pinned to hardware you already own.

How to decide

If you are building a game, a training environment for agents that must operate in a shared space, or anything where the world has to stay consistent while multiple actors move through it, Agora-2 is the more relevant piece of research and the less available one. Read the report, play the preview, and treat the twenty participants as a statement about state design rather than a capacity you can rent.

If you need to evaluate physical outcomes — will this grasp hold, will this pour spill, does this plan survive contact with the scene — Agora-2 has nothing for you and PAN 2 is downloadable, documented, and honest about dropping a step in most plans. Budget the GPUs before you budget the engineering: four B300s for two minutes per thirteen seconds sets the cost of every candidate your planner considers.

The two will be compared for a while, because "world model" is currently a label applied to anything that predicts something visual. The label is doing no work here. One of these systems is a place that stays put while people move through it. The other is a question about what happens next, answered on demand, with the answer capped at twenty seconds. Confusing them costs you a quarter of engineering time; telling them apart costs an afternoon.