PAN 2: MBZUAI's open 29B world model is now in ComfyUI, with an int8 build for 32 GB cards
MBZUAI's Institute of Foundation Models released PAN 2, an Apache-2.0 'world model' that turns a scene plus described actions into up to 20 seconds of 720p-class video. We missed the Nov 10 launch, but ComfyUI support merged on Nov 15 and Comfy-Org published an int8 build, so it's now something you can run locally on a single high-end consumer card, slowly.
What it is. PAN 2 is pitched as a world model, not a video generator. You give it the current scene (text, optionally a starting image) and actions in plain language, and it simulates what happens next as video. Under the hood a language model first rewrites your request into a structured, time-stamped description of events, actors and camera moves, then a 29B-parameter diffusion transformer trained from scratch renders it through the Wan 2.2 VAE. Qwen3.5-9B is the frozen text encoder. Output per the model card: 1248×704 at 24 fps, about 10 s by default and up to 20 s, text-to-video or image-to-video, no audio.
The launch. IFM announced it at 10:00 AM MST Nov 10 ('A language model predicts the next word. PAN 2 predicts the next world'), and the post passed 2 million impressions. Weights, training code, a Diffusers pipeline and a vLLM recipe are all public under Apache 2.0. The official repo ships a prompt-rewriter skill for the reasoning step, and the docs say long structured prompts work best.
In ComfyUI. Kijai's PR adding native PAN2 support merged at 12:15 PM MST Nov 15 (it's in ComfyUI master, after the v0.45.0 release). Comfy-Org repackaged the files the same day: a bf16 diffusion model, an int8 version, the int8 Qwen3.5-9B text encoder and the Wan 2.2 VAE. Image-to-video reuses the existing Wan 2.2 image-to-video latent node.
What it takes to run, per one early tester. @niklaslenz_ai ran the int8 build on an RTX 5090: about 14 minutes for a 10-second clip at 50 steps, about 40 minutes for 20 seconds from an image, with VRAM peaking at 31.4 of 31.8 GiB. The full bf16 transformer is around 63 GiB, so it won't fit a 32 GB card in one piece. He also notes the official safety guardrail isn't part of the ComfyUI path. Those are one person's numbers on one machine, not a benchmark.
Why video makers should care. Prompting with actions and their consequences ('the glass tips off the table, then...') is closer to how you'd block a scene than describing a single frame. Combined with long 20-second takes and a commercial-friendly license, PAN 2 is worth watching for previs and physical-action shots, even if render times rule it out for day-to-day work for now.
Signals: official launch post took off on Nov 10, plus ComfyUI native support and Comfy-Org weights Nov 15 and early tester posts. No GitHub Trending or tracked-curator pickup, so this isn't HOT by our rule.
Skills & repos in our directory
- PAN 2 · ComfyUI, Generation
- Comfy Previs · Repo, Generation
- seriesmith · ComfyUI, Generation
- ComfyUI Wan 3.0 templates · ComfyUI, Generation
- Comfy-Org workflow_templates · ComfyUI, Generation
- ComfyUI-WanVideoWrapper · ComfyUI, Generation
More from the Lab
- ComfyUI v0.46.0: a Video Deflicker node, 2.3x faster RAFT, mixed-frame-rate concatenation, and new Seedance, Kling and HeyGen partner nodes
- Apple FastFrames: open research video models that generate 480p frames in under 50 ms
- MiniMax H3-Image: open-weights 16 MP image model from the H3 family, with 16 references and day-one ComfyUI support
- HyperFlow H3 Flash: Video Rebirth's 14B, 6-step MiniMax-H3 that makes video and sound on one consumer GPU
Topics: ComfyUI · Open-weight video & audio models · Wan