MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI

MBZUAI’s Institute of Foundation Models just dropped PAN 2, a 29‑billion‑parameter diffusion transformer that promises to let robotics teams “test a plan in simulation before acting on it.” The hype train is full‑speed, but the reality under the hood is a mixed bag that every independent hosting provider and AI startup should size up before committing GPU racks to this beast.

What PAN 2 actually is

PAN 2 is a “general‑purpose world model” that takes a scene description plus a textual action and spits out a video simulation of the outcome. It follows the Generative Latent Prediction (GLP) architecture from IFM’s own paper and replaces the original PAN model released in November 2025. The model runs on data‑center GPUs, with each worker loading the full 29 B parameter model in BF16 and falling back to block offloading when needed. No VRAM or latency numbers are disclosed, but the requirement to hold the entire model per worker signals a hefty memory footprint.

The pipeline splits into two stages: a conceptual reasoner (any instruction‑following LLM, IFM uses Kimi K3 by default) that converts the text action into structured JSON, and a perceptual backbone—a 29 B MMDiT diffusion transformer—that predicts video latents. A frozen Qwen 3.5‑9B text encoder conditions the diffusion model, and the Wan 2.2 VAE decoder renders the final frames at 704×1248, 24 fps, up to 20 seconds (481 frames). The default clip is 241 frames, roughly ten seconds.

Performance on the numbers board

IFM’s self‑reported benchmark suite covers six capabilities across 300 single‑action cases and 100 multi‑step tasks. Step‑completion hovers at 75.1 %, essentially tied with MiniMax‑H3’s 75.12 %.

On the broader six‑capability metric PAN 2 scores 84.48 %, trailing MiniMax‑H3’s 87.20 %. MiniMax‑H3 also wins human Elo (1620 vs PAN 2’s 1595 with a GPT‑6 Astra reasoner, 1559 with Claude Opus) and leads in single‑action cases. Physics‑IQ verified scores place PAN 2 at 49.38, the top among open models but below closed‑source FLUX 3’s 51.11. RBench, a robotics benchmark, puts PAN 2 at 57.04, again just ahead of MiniMax‑H3’s 56.16.

The LLM dependency – a hidden cost

The ablation study is crystal clear: strip out the conceptual LLM and the model collapses. Full‑plan completion drops from 34.0 % to zero, and step‑completion plummets from 75.1 % to 8.7 %. A “format‑only rewrite” recovers some ground, lifting full‑plan completion to 27.8 %. This tells any founder that the LLM isn’t a nice‑to‑have add‑on; it’s the glue that makes the diffusion model useful for multi‑step reasoning.

IFM ships a “Conceptual Reasoner” skill for Claude Code and Codex that rewrites free‑text prompts into the structured JSON format the model expects. The token budget for captions is hard‑capped at 1,536 tokens, which can bite when you try to describe complex scenes. In practice, you’ll need to budget prompt engineering time and possibly run a separate LLM service alongside the diffusion model, adding latency and cost.

Hardware implications for independent providers

Running PAN 2 means each GPU worker must load the entire 29 B model in BF16. While IFM mentions block offloading as a fallback, the baseline memory requirement is massive. For providers with 40 GB‑class GPUs, you’ll be forced into multi‑GPU sharding or rely on off‑device paging, which throttles throughput. The default sampler runs 50 diffusion steps with flow shift 9 and guidance 2.5, a configuration that is compute‑heavy and will chew through GPU cycles quickly.

Because IFM does not publish generation latency, you’ll have to benchmark in‑house. Expect generation times in the low‑single‑digit seconds per ten‑second clip on high‑end A100‑class hardware, but that’s a guess. The lack of VRAM disclosure is a red flag for anyone budgeting GPU clusters; you’ll need to test on your own hardware before promising SLAs to customers.

Safety guardrails – a double‑edged sword

IFM bundles a “pan2‑guardrail” that slashes unsafe text‑to‑video (T2V) outputs from 33.9 % down to 1.5 % in their internal tests. It also blocks 11.5 % of benign requests, indicating a fairly aggressive safety filter. While this is good news for compliance, the false‑positive rate could frustrate end‑users who need legitimate content generation.

From a provider standpoint, you’ll have to decide whether to ship the guardrail as‑is or build your own moderation layer. The guardrail’s impact on throughput is unknown, but any extra inference step will add latency and cost. Keep an eye on how the filter behaves with domain‑specific prompts – you might need to tune thresholds.

Business‑risk lens – is PAN 2 worth the investment?

If your product hinges on high‑fidelity multi‑step simulation—say, robotic manipulation planning or virtual prototyping—PAN 2 offers the only open‑source model that currently leads on full‑plan completion. However, the edge over MiniMax‑H3 is modest (34 % vs 32.5 %), and MiniMax‑H3 still dominates single‑action quality and human preference.

For most SaaS AI startups, the hardware bill will dwarf any licensing cost, since the model and code are Apache 2.0. You’ll need to allocate GPU capacity for both the diffusion model and a companion LLM (Kimi K3 or an equivalent). If you already run large language models, you can piggy‑back the reasoner; otherwise, you’re adding a whole new stack.

Actionable advice for founders and hosting operators

1. **Benchmark on your own rig** – Pull the weights from Hugging Face, spin up the Diffusers or vLLM pipeline, and measure memory, latency, and cost per clip. Don’t rely on IFM’s undisclosed VRAM figures.

2. **Plan for dual‑model deployment** – Allocate separate GPU pools for the LLM reasoner and the 29 B diffusion model, or use a sharding framework that can keep both on the same hardware without thrashing.

3. **Factor safety filters into SLA** – The pan2‑guardrail cuts unsafe output dramatically but also blocks a chunk of benign requests. Test its false‑positive rate on your domain and decide whether to expose it to users.

4. **Consider alternatives for single‑step use cases** – If you only need short video generation or single‑action prediction, MiniMax‑H3 or other smaller models may give you better quality at lower cost.

5. **Watch the roadmap** – IFM dropped gameplay and animation footage after early training, suggesting they are still pruning the dataset. Future releases may tighten the model or shrink the footprint; stay tuned.

Bottom line: PAN 2 is a bold step toward open physical AI simulation, but the hardware appetite and LLM dependency mean it’s a heavyweight play. Independent providers can leverage it for niche, high‑value simulation workloads, but must do the math on GPU spend and latency before promising it to customers.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Reporting is based on the source material cited below. Sources: MarkTechPost; marktechpost.com; Global1.News (12 November 2026).

By Allan Ali, Global1.News