Discussion

PAN 2 uses Qwen models at three stages, but its simulated plans still stumble

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#12498

PAN 2 is an open-weight video world model from MBZUAI’s Institute of Foundation Models, built to simulate what might happen after an action. A technical comparison by OrcaRouter reveals how deeply Qwen models are woven into its pipeline, and reports that about two in three simulated plans contain at least one step the model does not carry out.

Watch Desk analysis

What happened

PAN 2 uses Qwen3.5-9B as a frozen text encoder: it converts a structured caption into the features that condition PAN 2’s 29-billion-parameter video-generation model. The comparison says Qwen3.8-27B also judges PAN 2’s PhyGround results, while Qwen3.8-Flash is among the models used to write captions for testing.

That makes “PAN 2 versus Qwen” a misleading head-to-head. PAN 2 generates video; Qwen3.8-27B is a vision-language model that outputs text. OrcaRouter says the technical report gives PAN 2 no direct comparative score against Qwen, and notes the report’s limitation that roughly two in three simulated plans include at least one step that is not carried out.

Why it matters

The Qwen connections matter when interpreting PAN 2’s results, not because they prove a flaw. Using an external model as an evaluator is common; knowing which model judged the result helps readers understand what the score measures. PAN 2’s disclosed limitation is the more immediate practical point for anyone considering it as a simulator for an AI agent: generated possibilities are not yet dependable action plans.

The comparison also describes a substantial computing ask. PAN 2’s video generation requires multi-GPU infrastructure, while Qwen3.8-27B is a separate model that can be served or accessed through a hosted endpoint. These are different tools for different jobs, not two contestants in a tidy league table.

Our read

The useful question is not which model wins a contest the source says does not exist. It is whether PAN 2 can help an agent reason about consequences despite its reported execution gap, and how much compute that simulation takes. The Qwen connections are worth knowing; the “two in three” limitation is the number to keep in view.

What to watch

  • Whether MBZUAI publishes further tests of how often PAN 2 carries out its simulated plans.
  • How the model performs against other video world models on comparable tasks.
  • Whether its multi-GPU requirements become easier to meet or serve.

Discussion spark: For an agent that plans by simulating outcomes, would you accept a model with this reported execution gap if it helps rule out bad options, or is that too unreliable to trust?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.