PAN 2 vs Genie 4: Both Put a Language Model in the Loop, and Only One Names It

The two most interesting world models of late 2026 are built the same way, and neither company leads with it. PAN 2 is the Institute of Foundation Models' 29-billion-parameter open-weight world model, published as an Apache-2.0 checkpoint on 7 October 2026 with inference and training code following on 5 November and Day-0 serving support in vLLM-Omni on 10 November. Genie 4 is G​oogle DeepMind's real-time interactive world model, announced on 27 October 2026, which generates an explorable scene from a text description and lets you steer it with a keyboard while native audio plays along. Put them side by side and the expected contrast — open weights against a research preview, twenty seconds of video against a live session — is real but boring. The interesting one is that both delegate the thinking to a language model, and only PAN 2 tells you which language model and lets you change it.

That single difference in disclosure is worth more than a benchmark table, because it decides whether you can reason about the system at all. A world model with an unnamed reasoning half is a product. A world model with a named, swappable reasoning half is an architecture.

The shared shape: a reasoner writes the future, a renderer draws it

PAN 2 implements Generative Latent Prediction, the pattern formalised in Critique of World Models (arXiv 2507.05169). A conceptual reasoner — any instruction-following language model — reads the current observation and the requested actions and writes the next world state as a structured, time-grounded description: events in order, each action bound to the actor performing it, plus camera motion. A frozen Qwen3.5-9B text encoder turns that description into conditioning features. A 29B multi-modal diffusion transformer trained from scratch on hundreds of millions of annotated videos fills in the perceptual detail the description leaves open, and a Wan 2.2 VAE decodes the result into frames.

G​oogle's published description of Genie 4 has the same division of labour wearing different product names. G​emini Omni grounds the simulation in realistic appearance and behaviour; other G​emini models supply reasoning for coherence and continuity; and Gemini 3.8 Flash acts as a guiding agent that observes the scene, plans steps and checks progress. That last role is an in-loop verifier — a language model watching the world model and keeping it on task. G​oogle also paired Genie 4 with Gemini 4 Argon for multi-step robotics simulation, including a demonstration of an Apptronik Apollo humanoid playing checkers.

So in both stacks the pixels are the cheap, fast half and the language model is the half that decides what should happen. This is the shape the field converged on in 2026 without announcing it: a world model is not a video generator with a physics prior, it is a planner plus a renderer, and the planner is an ordinary text model you could meter.

That convergence is why the disclosure gap matters. IFM ranked four planners by how often a simulated plan completes end to end — Claude Opus 5.5 first, GPT-6 Astra at medium effort second, Kimi K3 third, Qwen3.8-Flash fourth — and published the ranking along with the harness that produced it, while stating that the conceptual reasoner is by definition interchangeable. G​oogle named Gemini 3.8 Flash in the loop, which is more than most labs say, and published nothing about how you would substitute it, price it separately, or measure whether a different model would plan better.

Twenty seconds in a batch, or a session you steer

The interaction models are opposites, and this is where the two stop being comparable and start being different tools.

• Output — PAN 2 returns up to twenty seconds of 704×1248 video at 24 frames per second per call; Genie 4 returns a live, persistent scene whose resolution, frame rate and duration G​oogle has not published.

• Control — PAN 2 takes one starting frame plus a structured action caption refined by a prompt-rewriter skill; Genie 4 takes a text prompt and then continuous keyboard input for movement and camera.

• Audio — PAN 2 has none; Genie 4 generates ambient sound, music-like audio and interactive character dialogue alongside the visuals, which G​oogle describes as a first for a real-time world model.

• Batch behaviour — PAN 2 is a batch oracle: one call, one clip, no feedback inside the call; Genie 4 is a closed loop where your input at second three changes second four.

• Cost — PAN 2's documented default is eight H200s in BF16, with a vLLM recipe putting thirteen seconds of 704×1248 video at roughly 119 seconds on four B300s; Genie 4 has no price because there is no self-serve access.

• Access — PAN 2's weights are 86.17 gigabytes across sixty files under Apache-2.0, ungated; Genie 4 is a research preview for trusted testers with a stated path to AI Ultra subscribers through a waitlist and to early-access partners through the G​emini API and G​oogle AI Studio.

Read that list and the practical answer splits cleanly. If you want to ask "what happens if" a few thousand times and score the answers, PAN 2 is the only one of the two that can be called in a loop you control. If you want to explore, steer and listen, Genie 4 is the only one of the two that offers a place at all — and you probably cannot get in yet.

The specification sheet G​oogle did not publish

The most striking thing about Genie 4's announcement is the length of the list it leaves out: no resolution, no frame rate, no parameter count, no price, no measured scene-consistency duration, no benchmark with named baselines. What G​oogle did publish is a set of admissions — coherence holds over brief sessions, and visual details and object permanence can drift during longer runs, after complex edits, or while navigating. Complex physical interactions, audio accuracy, legible in-world text, spatial accuracy and long-term consistency are listed as open research problems. Every output carries a SynthID watermark.

None of that is evasive. It is the honest shape of a research preview, and listing drift as an open problem is more than most launch posts do. But it means a reader cannot compute anything from Genie 4: not clips per dollar, not seconds per GPU, not the duration at which a scene stops being usable. The one number a buyer would most want — how long a session stays coherent — is the one the announcement describes qualitatively and does not quantify.

PAN 2 answers the same question with a number that is worse and therefore more useful. Its model card states that about two in three simulated plans contain at least one step that is not carried out. It says appearance drifts across clips in long interactions because each clip is simulated from at most one starting frame plus text, and names perceptual memory as the missing piece rather than claiming it is solved. It concedes that text expresses "open a drawer" well and "place each finger here" badly, which bounds dexterous manipulation. It gives a concrete shortcut example — pasta appearing in the colander before it is drained — and admits that simulating plans which should fail is, in the card's words, not yet measured.

That last admission is the one that should follow any world model into a product spec. A simulator that can only render success is a generator of plausible outcomes, not a filter for bad ones, and the entire premise of simulative reasoning — simulate several candidate plans, compare, act on the best — depends on the model being able to show you a plan failing. PAN 2 says out loud that it has not been tested on that. Genie 4 does not say either way.

Failover is not a feature of either model, and it is the feature you need

Both of these systems are research artefacts pointed at a problem that will be solved by teams with production deadlines, and the gap between those two things is where a routing layer earns its place. The renderer half of a PAN 2 stack is your hardware and always will be — OrcaRouter does not host PAN 2 and cannot route it, and no catalogue can route a 29-billion-parameter diffusion transformer you are running on your own node. What is routable is the half that decides what to simulate, because that half is ordinary metered text work.

All four planners IFM ranked are in the OrcaRouter catalogue: Claude Opus 5.5, GPT-6 Astra, Kimi K3 and Qwen3.8-Flash. That makes the reasoner slot the one part of a PAN 2 pipeline you can benchmark this afternoon without renting a GPU node — point the prompt-rewriter at each of them through a single key, hold the diffusion half fixed, and count how often your plans survive. Provider list pricing passes through at 0% markup, so the cost comparison is the vendors' own rates rather than ours, and automatic failover absorbs a provider wobble mid-sweep instead of aborting the run.

The same route covers the Genie 4 side, narrowly. Gemini 3.8 Flash is the model G​oogle describes as Genie 4's in-loop guiding agent, and it is callable through the same catalogue. You cannot get Genie 4. You can build the part of the architecture that G​oogle did describe — a vision-capable model watching generated frames, planning the next steps and checking progress — and point it at a world model you do have. That is not the same product. It is the same control layer, and it is the layer where most of the engineering actually lives.

Be clear about what the ranking above is worth: IFM's ordering was measured on IFM's tasks with IFM's harness, against IFM's own baselines, and nobody has reproduced it. It is a shortlist, not a leaderboard. Running your own prompts through four candidates is a couple of hours of work and the only evidence that applies to your problem.

The other number worth carrying into any budget is the serving one. Fifty sampler steps at a flow shift of 9 with guidance of 2.5 on every step means a hundred forward passes through a 29B transformer per clip, which is why guidance has to be parallelised across ranks rather than run sequentially. A planner that evaluates five candidate actions per step over a ten-step task is asking for fifty clips, which at the published four-B300 configuration is a node busy for the better part of two hours before anything has been learned from the output. Simulative reasoning is not expensive because the model is large. It is expensive because the idea is a loop.

What each one refuses to say about duration, drift, and re-running a clip

Five dimensions decide whether either of these is usable in a pipeline, and none of them is a quality score.

• Session length — PAN 2 caps a call at twenty seconds and publishes that cap; Genie 4 runs a live session of unpublished length and describes coherence as holding only over brief sessions.

• Drift — PAN 2 says appearance drifts across clips because each one starts from at most one frame plus text, and names perceptual memory as the missing piece; Genie 4 says detail and object permanence can drift during longer runs, after complex edits, or while navigating.

• Re-running a clip — PAN 2 documents its sampler, including a default seed of 123 and a config file written beside every video; Genie 4 publishes no seed control, so a session is not something you can re-execute.

• Cost per hypothesis — PAN 2's cost is computable before you spend it: the published recipe puts thirteen seconds of 704×1248 video at roughly 119 seconds on four B300s; Genie 4 costs nothing to try and cannot be budgeted, because there is no price to multiply.

• What survives the run — PAN 2 leaves an MP4, the config that produced it and the caption you supplied; Genie 4 leaves a recorded session carrying a SynthID watermark and no artefact you can re-execute.

Duration is the dimension that decides most real decisions, and it is the one where the two are least comparable. PAN 2 gives you exactly twenty seconds and tells you it drifts across clips because each is simulated from at most one starting frame plus text. Genie 4 gives you a session of unspecified length, admits that coherence holds only over brief sessions, and calls long-term consistency an open research problem. Neither will tell you the duration at which the output stops being trustworthy, and that number — not a benchmark score — is what a pipeline actually needs.

What would make this a real comparison

One thing: a shared task. Nobody has run PAN 2 and Genie 4 on the same question, and it is not obvious they could — one produces fixed-length clips from a caption, the other produces an interactive session from a prompt and a keyboard. The nearest thing to a common measure would be action adherence under a described plan, and only PAN 2 has published a harness for that: 300 single-action cases across six capability categories, 100 multi-step tasks posed as 200 prompts and 456 steps, scored on step completion and full-plan completion over three seeds. Thirteen reviewers produced the human-preference numbers on 140 prompts in a blind pairwise study, judging the authors' model against the authors' baselines. That is not independent. It is legible, which is the next best thing, and it is more than the other side offers.

So the choice, stated as plainly as the evidence allows. If you need a world model you can download, audit, fine-tune and query in a batch — and you own or can rent a multi-GPU node — PAN 2 is the most complete open release in the category, and its own limitations section is the most valuable page in it. If you need a world a person can walk around inside, with sound, and you are willing to wait for a waitlist and build against an API that has no date and no price, Genie 4 is the more impressive system and the one with a distribution path behind it.

The mismatch to avoid is treating either as a substitute for the other. Genie 4's breadth is not evidence it can grade a plan; PAN 2's downloadable weights are not evidence it can hold a scene together while you steer. What they genuinely share is the architecture both companies arrived at independently — a language model deciding, a diffusion model drawing — and the fact that in 2026 you can still only see inside one of the two.