Is PAN 2 state of the art?Pending benchmarks.Announced Nov 10, 2026; no independent results yet.

The details so far

Lab

IFM (MBZUAI) with Inception42

Announced

Nov 10, 2026

API

Not yet · No public cloud API announced; run locally from the released weights and code

Open weights

Published · Public, ungated Hugging Face repo IFM/PAN2 (BF16 diffusion backbone). The Qwen3.5-9B text encoder and Wan 2.2 VAE are downloaded from their own repositories

Parameters

29B-parameter MMDiT perceptual reasoning backbone, trained from scratch. Conceptual reasoning is a pre-trained multimodal LLM (Kimi K3 is used in the technical report unless stated otherwise). Frozen text encoder is Qwen3.5-9B; VAE is Wan 2.2, both used unmodified from their authors

Type

General-purpose world model with Generative Latent Prediction (GLP): conceptual reasoning by a multimodal LLM plus perceptual reasoning by a 29B latent diffusion model (IFM page, model card and GitHub)

Output

Video via the Wan 2.2 VAE decoder, 704×1248 at 24 fps, up to 20 seconds; text-to-video and image-to-video from one diffusion checkpoint (model card)

License

PAN 2 code and weights Apache License 2.0 (IFM announcement, model card and GitHub). Third-party Wan 2.2, Qwen3.5-9B and Diffusers components keep their own licenses

Code

https://github.com/ifm-ai/PAN2 (training and inference code; also Diffusers, vLLM and FastVideo)

Related

Separate from earlier PAN, from Jais 3 (same lab), and from the frozen Wan 2.2 VAE and Qwen3.5-9B encoder used as components

Tags marked unsure are not stated on the lab's own pages.

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

• Publicly announced with a source
• Usable by the public (API or weights)
• Listed on an independent board we track
• Enough results for a rank in a category
• Holds across two snapshots (crowns only)

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: on 100 egocentric everyday multi-step tasks posed as I2V and T2V (200 prompts, 456 steps), with Kimi K3 as the conceptual reasoner, PAN 2 records 75.10 ± 1.51% step completion and 34.00 ± 3.61% full-plan completion, mean ± SD over 3 seeds (IFM technical report figure on the Hugging Face model card).

Lab claim: six-capability overall accuracy of 84.48 ± 0.08% (physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint control and external dynamism; 300 single-action test cases; IFM technical report figure on the Hugging Face model card).

Lab claim: Physics-IQ Verified (I2V) 49.38 ± 0.64 in the public leaderboard setting, with Claude Opus 5.5 as the conceptual reasoner, mean ± SD over seeds 123–126 (IFM technical report figure on the Hugging Face model card).

Sources