Qwen’s World Model Outperforms Claude Opus 4.8

Qwen recently released a language world model (LWM) that learns to simulate the environments agents act in rather than to act in them, which is similar to dreaming techniques in robotic applications (e.g., DreamGen, Cosmos).

A world model predicts environment dynamics based on current observations and actions, serving as a core cognitive mechanism for reasoning and planning.

Qwen’s world model is a 397-billion-parameter model that out-simulates GPT-5.4 on a new benchmark.

The benchmark win is noise here because 397B model that produced it scores 58.71 on AgentWorldBench versus 58.25 for GPT-5.4, a 0.46-point margin on a metric scored by an LLM judge. (more on this later)

And that model isn’t even the thing you can download because the open-weight release is the smaller Qwen-AgentWorld-35B-A3B, which scores 56.39, below Claude Opus 4.8 (56.59), but higher than Claude Sonnet 4.6 (56.04).

So let’s not read the marketing but truly understand why does this is very important:

1. Agents trained inside deliberately fictional environments that transfer to real tasks.
2. A controllable simulator that injects targeted perturbations trains better agents than the real environment does, on some axes by a wide margin.
3. Training a model to predict environments improves it at acting in environments, with zero agentic fine-tuning, including on tasks the world-model training…