PAN 2 Opens Its Weights, and a Single Ablation Explains the Whole Design

On 10 November, the Institute of Foundation Models at MBZUAI released PAN 2 with weights, training code and inference code under Apache 2.0. Few world models ship their training code, and the press release leans on that. The technical blog is the part to read, because it contains one ablation that reframes what the system is.
Give PAN 2 a starting scene and an action described in words, and it outputs up to twenty seconds of video showing what happens next. The claim is not that the video looks good. The claim is that the sequence is a plausible chain of physical consequences, which is a different and much harder thing to measure.
Two reasoners, one checkpoint
The architecture is what IFM calls Generative Latent Prediction. It splits the job in two.
A multimodal language model reads the starting image and the requested action, then writes a structured, time-stamped description of the next world state. Who acts. When. How the camera moves. That description is JSON, not prose, which matters because it becomes an interface you can inspect.
A 29B parameter multimodal diffusion transformer, trained from scratch on real footage chosen for physical dynamics rather than visual appeal, then fills in what language leaves open. Contact. Deformation. The way a wave breaks. A frozen Qwen3.5-9B text encoder conditions it, and a Wan 2.2 VAE decodes frames.
The released checkpoint handles both text-to-video and image-to-video. Defaults are 704 by 1248, 24 frames per second, 50 sampling steps, CFG 2.5, and up to 20 seconds. Camera motion is described separately from actor motion, which is how the scene stays coherent when the viewpoint changes.
IFM's default planner is Kimi K3. In the leaderboard configuration on Physics-IQ Verified they used Claude Opus 5.5 as the conceptual reasoner. That detail is worth holding onto, because it means the headline score is partly a statement about someone else's language model.
The ablation is the story
Here is the table that should get quoted more than the benchmark rank.
Remove the conceptual reasoning step and feed the action straight to the diffusion model, and multi-step plan completion drops from 34.0 percent to 0.0 percent. Single-step completion drops from 75.1 percent to 8.7 percent.
Rewrite the prompt into the same structured format but keep only surface formatting, and full-plan completion recovers to 27.8 percent. So most of the value sits in the reasoning, not in the formatting, and a little of it sits in the struct.
Read that again in plain terms. Without a planner that decides who does what and when, the video model cannot finish a multi-step task at all. Not badly. Zero. The thing being sold as a video generator is closer to a renderer bolted onto a scheduler, and the scheduler is where the physical reasoning lives.

That framing also explains the failure mode IFM admits to. About two out of three simulated plans still miss at least one step. One in three lands. And the model card notes a specific pattern: the simulation sometimes skips a necessary physical step while still moving toward the requested outcome. A convincing final frame can hide a broken sequence. If you only look at the last second of the clip, you will believe it.
Where it sits
Multi-step plans across 100 everyday tasks, full-plan completion:
| Model | Full plan completion | | --- | --- | | PAN 2 | 34.00% | | MiniMax-H3 | 32.50% | | FLUX 3 | 26.77% | | Seedance 2.5 | 24.49% |
On the six dimensions IFM uses for physical intelligence, physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamics, PAN 2 scores 84.48 overall. MiniMax-H3 scores 87.20. So PAN 2 leads on finishing a plan and trails on the individual capabilities inside it, which is a strange and specific shape for a result. It is better at keeping a sequence on track than at any single frame within it.
Physics-IQ Verified puts PAN 2 at 49.38 with Opus 5.5 as the planner. That is the best open-weight number, ahead of Cosmos 3 Super at 48.62 and behind the closed FLUX 3 at 51.11. PhyGround is 4.74, RBench 57.04. In a blinded human comparison on action adherence, physical plausibility and world consistency, PAN 2 ranked second among the four open models tested.
All of those figures come from IFM's own report. The public leaderboard setting is reported by the authors, which is the usual arrangement and the usual caveat.
Installing it
Weights are at IFM/PAN2 on Hugging Face. Code is at ifm-ai/PAN2, with a separate ifm-ai/pan2-guardrail repository for the safety filter, which IFM strongly recommends running alongside. Day-zero support exists in Diffusers and vLLM-Omni.
The compute bar is high and the docs do not hide it. Each inference worker loads the full 29B model and takes a share of the prompts, so extra GPUs buy concurrency rather than speed on a single request. Offloading options move blocks between CPU memory and the GPU, and how much that helps depends entirely on your hardware.
There is a workflow detail that is easy to miss and important. IFM ships a Conceptual Reasoner skill that works in Codex or Claude Code, and the recommendation is to rewrite free-text prompts into structured captions with it before generation. That is the ablation showing up as a product decision. Prompt engineering here means producing a usable schedule rather than choosing better adjectives.
A sane first trial is small and fixed. Save the starting image, the rewritten prompt, the config and the output together, so that when a sequence goes wrong you can tell whether the failure came from an ambiguous action description, a structured prompt that dropped a condition, or the video stage losing an object.
What this is for
IFM is explicit that this is a foundation for agents, not a toy for demo clips. The intended loop is: an agent proposes an action, PAN 2 simulates the consequence, the agent compares outcomes and revises, and new observations feed back in. For that loop to be worth running, the simulation has to be checkable against reality rather than merely convincing.
That is where the 34 percent matters. If one in three plans completes and two in three drop a step, an agent using this to plan is going to be wrong often enough that you need a verifier in the loop, and possibly a human. The team says as much. Long-term world consistency remains open, prompt sensitivity is documented, and consistency across clips is documented as a problem.
Two things make the download worthwhile anyway. The weights and the training code let you reproduce the failures instead of taking them on trust. And the planner ablation gives you a clean, cheap experiment: swap the language model, keep everything else, and watch whether plan completion moves. That is a much better use of an afternoon than rerunning a leaderboard.
What PAN 2 actually asks is where physical reasoning should live. On Physics-IQ it trails the closed FLUX 3 by a couple of points; on full plans it leads every model tested, closed ones included. Those two results only fit together if the reasoning happens in the planner and the diffusion backbone does the rendering. PAN 2's own numbers point at the language model, and at a video model closer to a very good decoder than to a mind.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.