PAN 2 vs SKY-001: One Simulates What Happens Next, One Draws What to Build

Kimi K3 appears in both launch documents, in the same fortnight, doing opposite jobs. In PAN 2's technical report it is the conceptual reasoner — the language model that reads a request and writes the structured, time-grounded caption that a 29-billion-parameter diffusion transformer turns into video. In Skyfall's SKY-001 tables it is a baseline competitor, asked to draw a CAD assembly itself, where it reproduces 9.5% of the reference geometry by intersection-over-union. One launch uses a frontier language model to write the plan. The other measures frontier language models failing at the drawing. That split is a better guide to these two models than the label they share. PAN 2 is the Institute of Foundation Models' world model at MBZUAI, Apache-2.0 weights on Hugging Face since 7 October 2026, code since 5 November and a first-party announcement on 10 November, and it predicts up to twenty seconds of 704×1248 video from an optional first frame plus an action described in text. SKY-001 is Skyfall AI's engineering model, announced 19 October 2026, still behind a waitlist with commercial deployment targeted for early 2027, and it predicts the ordered sequence of drawing operations that produces a CAD part or a floor plan.

Neither date is today, and nothing below is framed as a launch. What follows is what each model actually computes, where each one is weak by its own numbers, and which of the two your problem wants.

Two different things are being predicted

PAN 2's chain has a language model in it. A conceptual reasoner — any instruction-following model — reads the observation and the requested actions and writes the next world state as a structured caption: events in order, actions bound to actors, camera motion. A frozen Qwen3.5-9B text encoder maps that caption into conditioning features. A 29B multi-modal diffusion transformer, trained from scratch on video annotated in the same format, predicts the latents the caption leaves open, conditioned on it and on the VAE latent of the starting frame, and a Wan 2.2 decoder renders them. The default sampler is 50 steps at flow shift 9 with classifier-free guidance 2.5 on every step, and the checkpoint is 86.17 GB across sixty files.

SKY-001's chain has a hierarchy in it instead. A Hierarchical Composer takes a design intent — an image, text, or a building envelope with a brief — generates the whole design, then splits it level by level and refines each leaf in parallel. A Construction World Model then predicts the ordered sequence of visual states that builds each leaf: in CAD, a chain of 3D solid states with the 2D sketch drawn between each consecutive pair; in BIM, a floor plan drawn stroke by stroke. The drawing runs as diffusion in two passes, a coarse rollout in a compressed latent space from a Matryoshka-style autoencoder and then a refiner that restores full resolution. Both stages are diffusion transformers; neither predicts tokens. Skyfall has not published parameter counts, weights, a licence or an API.

• What is predicted — PAN 2 predicts the next observation as pixels, given an action; SKY-001 predicts the next construction state as a solid or a stroke, given an intent.

• How time enters — PAN 2 runs on a physical clock, 241 frames (about ten seconds) by default and up to 481 (twenty seconds) at 24 fps; SKY-001 runs on a procedural order of roughly 100 drawing steps per CAD component and 70 per floor plan, with no physical clock anywhere in it.

• Whether physics is claimed — PAN 2 is scored on physical plausibility directly; SKY-001's materials make no dynamics claim at all, because what it draws is a design, not a motion.

• What you get out — an MP4 you can watch, versus geometry you score by IoU against a reference shape.

• Weights — an Apache-2.0 download for PAN 2; nothing published for SKY-001.

• Price — GPU time you can size for PAN 2; a waitlist for SKY-001.

Both are two-stage, and the bottleneck sits in different places

The shared engineering bet is that you split the job. One stage commits to structure — an ordered description of what should happen — and a second stage fills in the detail that description leaves open. PAN 2 does it with a caption and a diffusion backbone. SKY-001 does it with a hierarchy of subgoals and a construction model per leaf. Put the two launches side by side and the interesting question stops being which architecture is better and becomes which stage is actually costing you.

For PAN 2 that is the planner, and the report says so twice. On Physics-IQ Verified, an image-to-video benchmark normalized so that the spread between two real takes of the same scene scores 100, PAN 2 reaches 93.2 ± 0.8 with Claude Opus 5.5 writing the captions and 90.9 ± 0.4 on its system default, over seeds 123 to 126. Swapping the language model that writes the caption moves the score 2.3 points on a hundred-point scale where a perfect score means indistinguishable from real footage. The same pattern shows in the model card's limitations: about two in three simulated plans contain at least one step that is not carried out, the agent case studies relied on hand-refined prompts, dexterous actions such as where each finger meets an object are specified only coarsely in text, and captions are validated against a soft limit of 1,480 and a hard capacity of 1,536 tokens. The renderer is not the reported problem. The sentence describing the scene is.

For SKY-001 the planner is not measured as a separate stage, so the reported bottleneck is the renderer. On Skyfall's CAD construction benchmark — 1,000 components sampled from objects such as a Jaguar E-Type Roadster, a FANUC ArcMate 100iB welding arm and a Bell 222 helicopter, each taking roughly 100 drawing steps — SKY-001 reproduces 79.7% of the reference geometry by mean IoU, against 79.1% for GPT-6 Astra and 76.9% for Gemini 4. On the BIM benchmark, 1,000 floor plans at about 70 drawing steps each, it reaches 43.6% against 42.3% for Gemini 3.8 Flash and 40.6% for GPT-6 Astra. Six-tenths of a point and 1.3 points are the architectural leads. Less than half a floor plan is the absolute number.

Read together, and read only for relative magnitude inside each vendor's own reporting, the planner is the bigger lever in either document. Changing PAN 2's caption writer moves it further than SKY-001's whole architectural advantage over its nearest rival. The benchmarks are different and the scores are not comparable; the point is where each team's own variance lives.

The benchmarks share no harness, no judge, and no baseline policy

PAN 2 is reported four ways, and three of them involve somebody else's judgement. Physics-IQ Verified is a public benchmark with a leaderboard setting, and PAN 2's entry uses Claude Opus 5.5 as the reasoner while the system-default column runs each competitor's own pipeline as IFM drove it. PhyGround is a five-point scale judged by Qwen3.8-27B, which is to say another model's opinion rather than a physical measurement; PAN 2 scores 3.98 and 4.11 on the two configurations the report reports. RBench is an unweighted mean over five task and four embodiment categories multiplied by 100, where PAN 2 scores 83.7 and 73.5. The closed-model baselines carry a dagger because they were generated once each rather than over seeds. There is also a blind human study: 140 prompts, 13 reviewers, 1,400 pairwise comparisons, fitted to a Bradley–Terry model with Elo centred at 1500 and 95% confidence intervals. And there is IFM's own harness — 300 single-action test cases split 201 image-to-video and 99 text-to-video across six capabilities, plus 100 multi-step egocentric tasks posed as 200 prompts and 456 steps, reported as mean ± standard deviation over three seeds.

SKY-001 is reported two ways, both Skyfall's, both scored by IoU against a reference shape, both with the rival models driven by Skyfall with one-shot prompting. There are no error bars on any of it, and no third party has reproduced a number. Reading the tables carefully turns up three things worth knowing before you quote them.

First, the vendor's own pages disagree with each other. The 3D CAD panel on the product page shows SKY-001 at 76.1% against GPT-6 Astra at 73.4%, with Claude Opus 5.5 at 61.8%, Gemini 3.8 Flash at 45.9%, Qwen 3.8 Max at 26.6% and Kimi K3 at 6.5%. The construction benchmark reports 79.7% against 79.1%. Different tasks, different scores, same model names on both — and Claude Opus 5.5, which is the reasoner that lifts PAN 2's Physics-IQ number, appears in only one of them.

Second, the speed multiplier is stated three ways across three documents published the same day: 51× in the blog chart, 49× on the product page, 100× in the press release. The arithmetic settles it — 199 seconds divided by 3.9 is 51.0 — but it tells you the headline numbers were written by different people who did not diff their work.

Third, the cost column is unexplained. Skyfall reports $0.005 per CAD case and $0.0024 per plan, and does not say whether that is a list price, rented GPU time, or amortised self-hosted inference. Those are three different claims, and the difference matters most for the models Skyfall ran itself rather than called.

None of that makes SKY-001's efficiency story wrong. A 51-fold speed advantage and a 150-fold cost advantage over the best rival, on accuracy that ties, is a real claim about a real workload. It makes it unverified, which is a different thing.

What it costs to run, and how much of that is knowable

PAN 2's compute is published as a recipe you can price yourself. The default deployment is eight H200s in BF16, with classifier-free guidance split across two ranks and Ulysses sequence parallelism across two, and the VAE decode patched across four tiles. The serving recipe reports thirteen seconds of 704×1248 video in roughly 119 seconds on four B300s. There is no per-second list price because IFM does not run a hosted endpoint; the number you care about is what your provider charges for the GPUs, multiplied by the seconds per clip you actually generate.

SKY-001's compute is not published in any form you can price. There is no parameter count, no hardware description, no throughput figure, and no endpoint. What exists is a vendor cost per case whose composition is undisclosed and a waitlist form.

That asymmetry decides more of this comparison than the benchmark gaps do. One of these models can be benchmarked on your own data this afternoon, on your own hardware, at a cost you can forecast. The other can be read about.

The one part of either stack you can rent today

Neither PAN 2 nor SKY-001 is on OrcaRouter, and neither can be routed through it: PAN 2 is an 86 GB checkpoint you deploy yourself, and SKY-001 has no API to call. The part of PAN 2's pipeline that is a normal hosted model is the first one — the conceptual reasoner that turns your request into a world-state caption. IFM evaluated four of them in order of full-plan completion: Claude Opus 5.5, GPT-6 Astra at medium effort, Kimi K3, and Qwen3.8-Flash, and notes that any instruction-following model can be used. Qwen3.8-Flash is in the OrcaRouter catalogue, which means the caption stage of a PAN 2 deployment can run on one key with provider list pricing passed through at 0% markup, and with automatic failover if a provider degrades in the middle of a batch of plans. That is worth configuring before the diffusion stage, not after: the planner is where the measured variance is, so it is the stage where a cheaper model and a second provider are worth testing against each other on your own prompts.

Which problem are you actually holding

If the thing you need is a prediction of what an action does to a scene, in pixels, over the next few seconds — for an agent that wants to imagine outcomes before choosing, or for a video pipeline where the physics has to hold — PAN 2 is the one of these two built for it, and you can have it now. Budget for the caption stage and for the guardrail: IFM ships a separate guardrail repository that screens the request with Qwen3Guard-Gen-4B before generation and the output with a Q16 frame classifier after it, and in its internal evaluation that pairing cut the share of harmful text-to-video requests producing unsafe video from 33.9% to 1.5% while also stopping 11.5% of benign text-to-video requests and 3.3% of benign image-to-video requests. A safety layer that costs you one benign request in nine is a real production cost, not a footnote.

If the thing you need is a design intent turned into CAD or BIM geometry, SKY-001 is the only one of the two aimed at it, and you cannot have it. The honest stand-in today is a frontier language model wired into your CAD toolchain — which is precisely the setup SKY-001's own tables measure, at 9.5% to 42.3% IoU depending on the model and the task. That is why the waitlist is doing real work rather than marketing work: the alternative is not a competitor, it is a worse number.

If what you need is a physics engine, neither. PAN 2's simulations are predictions, its card says so in as many words, and its own limitations list includes simulations that progress while violating physical constraints — pasta appearing in the colander before it is drained. SKY-001 draws designs and makes no dynamics claim anywhere. Calling either one a world model is accurate in the sense the two teams use the term, and misleading in the sense a game engine uses it.

The thing to watch is not the next benchmark. It is whether anyone outside these two companies reproduces a number: whether Skyfall publishes the IoU protocol and the cost accounting behind its per-case figures, and whether the performance gap between PAN 2's default pipeline and its best reasoner survives contact with prompts nobody hand-refined. Both are questions about the same stage — the one that decides what should happen, before anything is drawn.