Cody: —so the headline is they finally shipped a real world model for agents. Qwen-AgentWorld—two models, 35B with 3B active and a 397B beast with 17B active—both trained to predict what happens next in seven different agent environments.
Justy: Right—predict the world before you act in it. That’s the missing layer everyone’s been dancing around.
Cody: Exactly. And they’re not just talking vibes. They built AgentWorldBench from real agent trajectories across nine existing benchmarks—Terminal-Bench, OSWorld-Verified, Tool Decathlon—then evaluated the world model’s simulation quality with open-ended rubric judging and rule-based verifiers.
Justy: So they built the benchmark they’re evaluated on—which, fine, but also kinda the point of the paper, right?
Cody: Yeah—
Justy: —and they’re calling it a prerequisite for general agents. Richens et al. get cited for the ‘must learn a world model to generalize’ claim.
Cody: Mm-hm. So the mechanism’s the headline: three-stage pipeline. CPT injects world knowledge and state transition dynamics. SFT activates next-state-prediction reasoning. RL—hybrid rubric-and-rule rewards—sharpens fidelity.
Justy: So it’s not just a bigger model talking to itself—it’s trained to predict the next observation given the action.
Cody: Right. And they’re not just dumping it into a single domain. Seven environments—MCP, Search, Terminal, Software Engineering, Android, Web, OS—unified under one textual schema so the model can hop domains without relearning the interface.
Justy: Okay, but who actually ships this? Is this a research toy or something teams will run in prod?
Cody: That’s where the two paradigms come in. First: decoupled simulator. You can spin up thousands of simulated environments for agentic RL—edge cases you’d never hit in real life, partial returns that force multi-step retrieval, targeted perturbations to expose agent weaknesses.
Justy: So it’s a sandbox generator with dials.
Cody: Exactly. Turn-level scaling without spinning up sandboxes or VMs. High-value domains where real execution is impossible—irreversible ops, proprietary setups—suddenly become safe to explore.
Justy: And the second paradigm?
Cody: Unified agent foundation model. World-model training warms up the agent before downstream RL. The agent learns to predict future states, then uses that internal simulation to refine action selection. It’s like reflection, but future-facing.
Justy: So the agent gets a built-in crystal ball.
Cody: Right. And next-state prediction itself is a forcing function for reasoning, knowledge, instruction following, long-context handling—all the things that make a general agent general.
Justy: This is the part I keep coming back to—why hasn’t anyone else done this before? Because it sounds like the obvious missing piece.
Cody: Because it’s expensive to train and the benchmarks didn’t exist until now. They had to gather more than 10 million environment interaction trajectories across seven domains just to get the data.
Justy: Fair. Okay, real talk—this is a read-it, reacted-to-it episode on our part. We could be a version or a number behind, so treat the specifics as a starting point, not gospel.
Cody: Sure—take it with a grain of salt.
Justy: But the mechanism still clicks. CPT for world knowledge, SFT for next-state reasoning, RL with rubric-and-rule rewards for fidelity—that’s a real training stack, not a slide-deck promise.
Cody: Agreed. And the two use-cases—decoupled simulator and unified agent—give product teams a clear knob to turn. If you’re building agentic RL that needs edge-case coverage, you plug in the world model as a simulator. If you’re training a general agent, you warm it up with world-model training first.
Justy: So the Build Next check-in—model weights are up on Hugging Face. Qwen-AgentWorld-35B-A3B.
Cody: Right. That’s the practical release.
Justy: No datasets or extra tooling beyond the model card—so if you’re itching to kick the tires, that’s the place to start.
Cody: Though the paper itself is still fresh. Give it a few weeks to bake.
Justy: Speaking of baking—out here in LA the light’s still gold at six thirty. Must be nice.
Cody: —blue hour already in DC. We’re both turning into the people who get excited about cache boundaries and confidence thresholds.
Justy: It’s worse than coherent. It’s productive. I’ve become the person who reads a world-model paper and gets giddy about next-state-prediction loss curves.
Cody: And I’ve become the person who explains how the loss curve works while sipping lukewarm coffee at eleven PM.
Justy: Exactly. Alright, catch you on the flip—