Agora-2 vs Kimi K3: Twenty Participants in a Shared World, and the 1M-Token Model That Plays One Role in It

Strip the benchmarks out and there is one asymmetry that decides this comparison. Odyssey introduced Agora-2 on September 21, 2026 — a playable multi-agent world model that generates a shared, interactive environment for up to 20 humans and AI agents in real time, at 640×480, from a learned simulation plus a per-view renderer. Kimi K3, from Moonshot AI, is a 2.8-trillion-parameter open-weights model with a 1,048,576-token context window and a 44 on the Artificial Analysis Intelligence Index, released July 15 and downloadable since July 27. Both are open in the sense that matters to a research team — Kimi K3's weights are on Hugging Face under a modified MIT license, and Agora-2's technical report is published in full — and neither is open in the sense that matters to a buyer: Agora-2 has no weights, no API and no price, and Kimi K3's licence carries commercial restrictions. The useful question is not which is smarter. It is which of them can be part of a multi-agent system this quarter.

What is actually downloadable, and what that gets you

Kimi K3 is the more conventional object by a wide margin. A 2.8T-parameter MoE with a million-token context, text and image input, a top-level reasoning-effort control in place of sampling parameters, native tool calling, and an OpenAI-compatible surface. It is priced at $3.00/$15.00 per million tokens with a $0.30 cache-read rate, and it scores 44 on index v4.3.2 — third among open-weights models on that board, behind MiMo-V2.6-Pro's 46 and GLM-5.3's 45. On the same board it is scored at 85.0 on terminal-bench 2.1 and 59.5 on SciCode. If your multi-agent system needs a brain per agent, this is a brain you can rent cheaply or run yourself.

Agora-2 is a different kind of artefact. What Odyssey published is a 23-page technical report and a browser preview. The report describes an isometric action-game environment with explicit representations for entity identity, position, health, animation, death and respawn; a simulation model that predicts movement, combat and animation from entity histories, actions and local geometry; and a rendering model per participant that generates that participant's own view from a compressed visual representation of the shared state. Nothing in that description is a language model, and nothing in it is downloadable.

• What it is — Agora-2 a learned game engine rendering 640×480 per view vs Kimi K3 a 2.8T-parameter open-weights text model

• Independent score — Agora-2 none published vs Kimi K3 AA Intelligence Index 44 (v4.3.2), open-weights leader

• Context — Agora-2 shared state plus bounded per-view history vs Kimi K3 1,048,576 tokens

• Price — Agora-2 none published, browser preview only vs Kimi K3 $3.00/$15.00 per 1M, $0.30 cached

• Licence — Agora-2 report public, no weights released vs Kimi K3 modified MIT, weights on Hugging Face

• Inputs — Agora-2 browser controls plus 16 RL agent policies vs Kimi K3 text and image

The role each one plays in a multi-agent system

These two only meet inside a system that contains both. Kimi K3 is a candidate for the decision layer — the component reading tool output, writing code, and choosing the next action in a long-horizon loop. Its million-token window and $0.30 cached-read rate are aimed exactly at the workload agentic loops generate: enormous, repetitive contexts that would be ruinous at full input price.

Agora-2 is a candidate for the layer below — the environment. Odyssey's own framing is that multi-agent world models let researchers study how agents interact "within controlled simulations, where harmful behaviour can be investigated without exposing real-world systems to risk," a sentence that reads as a direct response to the METR report in which roughly 1,200 agents coordinated an intrusion from inside an evaluation sandbox. The policy that drives Agora-2's sixteen autonomous entities is not an LLM; it is a recurrent scalar and spatial policy trained with reinforcement learning, consuming a sixteen-observation history and emitting an action every four simulation ticks. If you want to know what a Kimi K3-class agent does when sixteen opponents are also making decisions, Agora-2 is one way to build the arena. It is not a way to replace the agent.

Reading the numbers on both sides

Kimi K3's 44 is measured by a party that does not work for Moonshot, on the same v4.3.2 index as every other model on this page, and it is the score that makes it a credible open-weights frontier model. Its context window, price and modality list are published facts.

Agora-2's figures come from Team Odyssey's own report. The headline result is a rendering comparison: a structured-prefix renderer reaching 30.11 dB PSNR against 20.67 dB for a VAE-based baseline on thirty monitoring clips with three sampling seeds each. The report is careful to say that this compares complete systems with unmatched training budgets and does not isolate architecture. The conditioning benchmark that shows a 45.8% denoiser-time reduction against cross-attention runs on randomly initialised denoisers with synthetic inputs — an execution-cost test, explicitly not an image-quality measure. The simulation evaluation covers single-step movement and animation prediction on held-out recorded examples.

And the limitations section says the rest. The 30-frames-per-second display target and the 15-latent-updates-per-second cadence are stated as targets; sustained frame rate and input-to-display latency during live multi-agent play have not been quantitatively evaluated. Long-horizon visual quality, cross-view agreement and end-to-end action adherence are unevaluated. Procedural layout variation exists within the training domain but transfer to new assets, rules or environments is untested. This is not a knock on Odyssey — the report is more candid about its own gaps than most launch material — but it is the correct prior for anything you read about Agora-2's capabilities.

Which one you can build on this week

If you need a frontier open-weights model in production, the answer is Kimi K3 and it is not close. Its independent 44, its million-token window, its image input and its $3/$15 rate with cheap cache reads are a deployable package that has been on the board since July. On OrcaRouter it is served at Moonshot's own list price with zero markup, which matters more than usual for this model: a long-context agent loop spends most of its tokens on repeated prefixes, and the $0.30 cache-read rate passed through unchanged is the difference between a fleet that is affordable and one that is not. Automatic failover across providers is the second half of that — the longer the loop, the more expensive a mid-run provider failure.

Agora-2 has no rate card, so there is nothing to route to and we do not host it. What it offers a multi-agent team is a research preview of an environment that can be played today in a browser, backed by a public architecture report, in a category that until now had no shared-world simulator outside a game engine. If your interest is in what agents do to each other, that is a genuinely useful thing to have and worth an afternoon. If your interest is in what agents can do, Kimi K3 is the model and the world model is not a substitute for it — the two sit at different layers of the same stack, and only one of those layers has a price you can put in a spreadsheet.