Agora-2: Odyssey's Playable World Model Runs 20 Participants on Four GPUs

Agora-2 puts twenty participants in one simulation and calls it a playable research preview. Odyssey published it on September 21, 2026 in a post titled "Introducing Agora-2: Advancing Multi-Agent World Simulation," signed by Oliver Cameron, and the number that travels is the twenty. The number that matters is the split inside it. A session supports up to four concurrent human controllers alongside sixteen autonomous entities, and those are not the same kind of participant — four people typing movement commands, sixteen policies executing them. Odyssey trained the policies itself, on stages of a game it also trained on: the paper says the world is learned from captures of Diablo II.

That distinction between controlled and simulated participants is where almost everything interesting about Agora-2 lives. It is also where most of the coverage will blur, because "20 players" is a cleaner sentence than "four players and sixteen learned agents sharing a state that survives off-screen." What Odyssey is shipping is not a video generator with multiplayer bolted on. It is closer to a game server whose physics, animation and rendering were all replaced by learned components, and whose only genuine content source is a dataset of recorded play.

Twenty participants is four people and sixteen policies

Odyssey is explicit that Agora-2 supports five times as many participants as Agora-1 and that it moved from simulating a single environment to simulating multiple environments with longer-horizon behaviour. The per-session shape looks like this:

• Human controllers — up to 4 concurrent, each supplying movement and action input

• Autonomous entities — 16 per session, driven by learned monster and companion policies rather than by players

• Total participants — 20 in one shared world state, which is the figure Odyssey leads with

• Environments — multiple, up from the single environment Agora-1 could simulate

• Content basis — captures of Diablo II, so the world, the assets and the rules are all inherited from one recorded corpus

• Release form — a playable research preview, not a product, with no weights, no licence and no price

The autonomous sixteen are not decoration. Odyssey's report describes training separate monster and companion policies, and the evaluation counts are specific: 23,771 final-stage monster-policy examples drawn from a base teacher plus recovery and retention data, and 128,005 examples for the companion policy, scored against 252 evaluation episodes and 24 evaluation cases respectively. Those policies are what make a four-person session feel populated. They are also the reason the participant count is a design choice rather than a capacity claim — sixteen is how many agents the world server was trained to carry alongside four humans, not how many it could theoretically hold.

The architecture is one small model and one large one

Agora-2 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 paper. The simulation model has 4,457,777 parameters — roughly 4.46 million, four layers wide, width 256, four heads, a four-step history window, fourteen movement branches and a separate head for facing. The renderer is a flow-matching transformer of about one billion parameters, sixteen blocks at width 2,048, five of them carrying causal temporal attention, run for five solver steps per update.

The renderer is conditioned on a 342-token prefix rather than on the raw world:

• Map — 144 tokens, a 12×12 local tile grid around the viewpoint

• Entities — 36 tokens, four for user-controlled participants and 32 for other entity slots

• Transient effects — 160 tokens for effects that are neither map nor entity

• View and roster — one token each for the camera action and the session roster

• Per update — 300 image tokens attending over those 342 conditioning tokens

• Codec — an RAE-based latent at 2 frames into 15×20×32, then x264 CRF 18 or NVENC QP 18 on the way out

Odyssey names the reason the conditioning is structured this way: a learned world needs to keep entity properties in a state that outlives any particular camera. The post puts it plainly — because entity properties are retained independently of any particular view, they remain available when an entity is out of frame. That is the sentence that separates Agora-2 from a video model. A generator that only conditions on recent frames loses the monster the moment you turn away from it; a world model that keeps an explicit state does not.

Four GPUs buy four players

The serving plan in the report is the least publicised part of Agora-2 and the most useful for anyone estimating what a world like this would cost. Odyssey's configuration 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 policy. The stated targets are fifteen latent updates per second with thirty displayed frames per second.

Two numbers in the appendix make that configuration concrete. At 165 watts, the team measured an inference reduction of 9.9% from a change to the renderer — 62.92 milliseconds down to 56.69 milliseconds over 3,200 calls per implementation. And the renderer's training was equally specific: sixty thousand updates on 4,232,000 segments followed by sixty thousand more on 4,332,800 segments, AdamW at a learning rate of 1e-4, batch 128, EMA 0.9999, run on 32 B200 GPUs. That corpus is itemised — 1,200,000 full-roster combat segments, 2,016,000 stratified special-ability segments, 504,000 corpse-traversal segments, 512,000 movement segments without autonomous entities, and 100,800 boss-portal lifecycle segments.

Read those two paragraphs together and the shape of the product is clear. The renderer is the expensive half by three orders of magnitude in parameters, one GPU per concurrent viewer in serving, and thirty-two B200s in training. The simulation model — the part that actually decides what happens in the world — is four and a half million parameters, which is smaller than most embedding tables. Agora-2 is a rendering problem wearing a game-engine costume.

The published numbers, and the one thing they do not show

The report's quantitative results are, by Odyssey's own framing, ablations of its own design choices rather than competitive benchmark scores, and they should be read that way:

• Structured prefix vs a VAE baseline — 0.001279 mean squared error and 30.11 dB PSNR against 0.010272 and 20.67 dB, a gain of 9.44 dB, over 90 paired evaluations from 30 clips and 3 seeds

• Renderer attention — 20.09 milliseconds per two-frame denoiser update with the structured prefix against 37.06 with cross-attention and 18.82 with pooled AdaLN, a 45.8% denoiser reduction and 34.1% on the core

• History dropout at 50% — 0.05695 streaming error and 0.19535 cold-start against 0.06041 and 0.21082 at no dropout, at slightly worse map-perturbation fidelity

• Simulation accuracy — 85.17% on movement branch prediction, 68.17% on the harder blocked-example subset, and 95.84% on animation mode, with predicted mode switches at 7.49% against 3.54% recorded

Every one of those figures is measured against another configuration of Agora-2. None of them is a comparison against a shipped system, and none of them describes live play. Section 8 of the report is candid about the gap. Transfer to new assets, rules or environments remains untested. Errors accumulate. Independently generated images can still disagree in appearance, visibility or event timing. And the sentence worth quoting in full: sustained frame rate and input-to-display latency during live multi-agent interaction have not been quantitatively evaluated. The fifteen updates per second and thirty frames per second above are targets, not measurements.

Where the preview lives, and what it does not include

The playable preview is at Odyssey's own site, reached through the demo address agora.odyssey.ml, which redirects to agora.odyssey.systems. The technical report is the PDF linked from the same announcement. What is not published is as notable as what is: no model weights, no repository, no licence, no inference API and no price. Odyssey describes the release as a research preview and treats the trajectory towards multi-agent interaction inside its foundational world models as future work, with PROWL listed as the research thread it is expanding and Odyssey-3 named as the eventual destination.

That matters for anyone planning to build against Agora-2, because the practical answer today is that you cannot — not through us and not through anyone else's hosted endpoint. OrcaRouter does not serve Agora-2, Agora-1 or Odyssey-3; those model routes do not exist in our catalogue. What our platform does carry is the rest of a build that might sit around a learned world: a single endpoint covering 200+ models, priced at each provider's list rate with zero markup, so when a vendor drops a price the change is live here the same day, and a routing layer that fails a request over automatically when a provider degrades. If you end up using a model to generate level layouts, write policy code or caption recorded play, that is the part we can actually help with.

What to watch

Agora-2 is a real result with a real caveat, and the two are easy to confuse. The result is that a shared, persistent, multi-environment world can be simulated and rendered for twenty participants out of a corpus of recorded play, and that Odyssey can point at the specific design decisions — explicit state, a structured conditioning prefix, a tiny simulation model — that made it work. The caveat is that the paper's own section 8 lists live latency, sustained frame rate, cross-environment transfer and error accumulation as unevaluated. Until someone publishes a frame-rate trace from a real four-player session, the honest summary is that Agora-2 proves the architecture and leaves the experience unmeasured.

The second thing to watch is the direction of travel. Odyssey's own framing puts Agora-2 as a step on the way to putting multi-agent interaction inside a foundational world model, with PROWL as the research expansion and Odyssey-3 as the named target. If that happens, the interesting question stops being whether a world model can carry twenty participants and becomes whether a world model can carry them into a world it was never trained on — which is precisely the transfer question the current report declines to answer.