MBZUAI's IFM Releases PAN 2, a 29B World Model Under Apache 2.0, With Self-Reported 34% Full Multi-Step Plan Completion
MarkTechPost reported on November 10, 2026, that MBZUAI's IFM released PAN 2, a 29B world model whose weights, training code and inference code are published under Apache 2.0 and can run via Diffusers, vLLM, FastVideo or repository scripts. IFM self-reports a 34.0% full multi-step plan completion rate, leading the 6 models it tested, but roughly two-thirds of plans still miss steps, and it trails MiniMax-H3 on single actions and human preference; VRAM and generation time were not disclosed.
What changed
IFM announced PAN 2 on X on November 10, 2026, and MarkTechPost reported it the same day. PAN 2 is a general-purpose world model for physical intelligence: it takes a scene and a textual action as input and outputs video of the scene after the action occurs. It follows the generative latent prediction architecture proposed in IFM's paper and is the successor to PAN, released in November 2025. The architecture has two stages. Conceptual reasoning uses an instruction-following LLM to write observations and actions into structured JSON (events, timed actions bound to actors, camera motion); IFM's report uses Kimi K3 by default. Perceptual reasoning uses a 29B dense MMDiT trained from scratch to predict video latents, conditioned by a frozen Qwen3.5-9B text encoder, with a Wan 2.2 VAE decoder rendering the final frames. Output specs are 704×1248 at 24 fps for up to 20 seconds, with a default clip of 241 frames, about 10 seconds, and sampling defaults of 50 steps, flow shift 9 and guidance 2.5. The report says weights, training code and inference code are all released under Apache 2.0 and can run through Hugging Face Diffusers, vLLM, FastVideo or the repository's own scripts, and it includes a conceptual-reasoner skill for Claude Code and Codex that rewrites free text into structured descriptions. IFM's ablation shows conceptual reasoning is indispensable: removing it drops full plan completion from 34.0% to 0.0% and step completion from 75.1% to 8.7%, while format rewriting alone recovers 27.8%. The report also says IFM removed game, animation and simulation material after early training, and a 9B test model trained without that material won 58% of votes in a blind test.
What it means for developers
For developers, separating availability from capability limits is more useful than reading the rankings. On licensing, the report says weights, training code and inference code are all Apache 2.0, a combination closer to a release you can build your own training pipeline on than open weights alone; but the full inference chain also depends on the frozen Qwen3.5-9B text encoder and the Wan 2.2 VAE decoder, and the report does not state the license terms for those two parts, so verify them before commercial use. On hardware, the report only says each GPU worker loads the full 29B model in BF16, with block offloading as a fallback, and gives no VRAM figures or generation times, so capacity planning depends on your own tests. On capability, these are IFM's self-reported numbers: on 100 multi-step tasks, full plan completion is 34.0% (MiniMax-H3 32.50%, FLUX 3 26.77%, Seedance 2.5 24.49%), and step completion of 75.10% is essentially level with H3's 75.12%; but the overall score on 6 single-action capabilities is 84.48%, below H3's 87.20%, with the report saying H3 leads in all 6 categories, and human Elo of 1595 is also below H3's 1620. In other words, if your workload is single-action video generation, this release gives no reason to switch models; if it is multi-step "simulate, then act" planning, it offers a self-hostable option that comes with training code. The ablation also shows it is not download-and-go: without conceptual reasoning, full plan completion falls to zero, and format rewriting alone only recovers 27.8%, so the engineering quality of the planner and prompts directly determines usability. The report also recommends deploying pan2-guardrail separately; in IFM's tests it cut unsafe outputs on harmful text-to-video requests from 33.9% to 1.5% while blocking 11.5% of benign requests, so guardrails reduce risk but carry a cost in false blocks.
Integration and evaluation
The following is conditional analysis and advice based on the report, not verified conclusions. The first decision is whether to build your own planner and prompt pipeline. The report says removing conceptual reasoning drops full plan completion from 34.0% to 0.0%, and format rewriting alone recovers 27.8%; if your workload is multi-step tasks, the practical choice is to bring the LLM planner and the structured JSON rewriting flow into your engineering scope rather than feeding free text straight to the model. This assumes the task really needs multi-step simulation; if you only generate single actions, the ablation conclusion does not apply. A testable measure is to compare free-text and rewritten structured-description inputs on your own multi-step task set for full plan completion and step completion, while noting that the description has a hard budget of 1536 tokens and whether rewriting gets truncated directly changes results. The second decision is whether to put the guardrail in the chain. The report says that in IFM's tests, pan2-guardrail cut unsafe outputs on harmful requests from 33.9% to 1.5% but blocked 11.5% of benign requests; if prompts come from uncontrolled users, leaning toward a default pre-filter is reasonable, while if use is internal and prompts are controlled, the cost of false blocks stands out more. A testable measure is to count both the benign false-block rate and the residual harmful-output rate on prompts you already have, setting an acceptable false-block ceiling before deciding to ship. The third decision is the serving stack and VRAM baseline. The report lists four paths (Diffusers, vLLM, FastVideo and repository scripts) and states the default of 241 frames and 50 steps, but gives no VRAM or generation times. The choice is to fix one path as a baseline, then measure each worker's peak VRAM, the per-clip time with block offloading on and off, and the resulting picture-quality differences. Since full plan completion is only about one-third, cost should be counted per successfully completed multi-step plan rather than per video, so that retry overhead from failed steps is included.
Limits and open questions
All scores in the report are self-reported by IFM and run on its own test suite, not reproduced by a third party. The same material is also internally inconsistent: the summary gives Physics-IQ defaults of 37.48 and 42.91 for MiniMax-H3, while a table gives 49.38 for PAN 2 and 39.79 for H3 under leaderboard settings; the two sets cannot be mixed, and the model card and blog also differ on plan scores for Cosmos 3 Super and LTX-2.5. Several points are unstated: VRAM and generation times are missing, so capacity planning needs your own tests; license terms for the Qwen3.5-9B encoder and Wan 2.2 VAE decoder used in the pipeline are not described; the 11.5% benign false-block rate comes from IFM's test set and may not reproduce on your prompt distribution; and whether the 20-second limit and 704×1248 output meet your simulation needs also requires testing. PAN 2 is not currently listed in the KimiSeek catalog, the report mentions no hosted API, and the weight and code channels both point to the vendor. The report notes that the Kimi K3 planner IFM uses by default has an entry in the KimiSeek catalog, but the planner is replaceable, and access methods and prices should follow the site's pricing page.