PAN 2 vs MiniCPM5-2B: One Ships Its Data, the Other Ships Its Prompts

OpenBMB released MiniCPM5-2B on 6 September 2026 and, alongside the 2.52-billion-parameter checkpoint, published the training corpora: UltraX-Preview for web pre-training, UltraData-Code with its tiered L0-to-L3 management, UltraData-SFT-Agent-2609 with half a million agent samples, and UltraData-RL-2609 with more than eighty thousand reinforcement-learning samples. That is an unusual thing to hand over, and it is the right lens for this comparison. PAN 2, published quietly by MBZUAI's Institute of Foundation Models with no launch post and no hosted endpoint, does something adjacent but not the same: it hands over its prompts. The repository ships a conceptual-reasoner skill with a system prompt, a fixed negative prompt and a token-budget validator for the structured world-state captions its diffusion backbone expects, plus the training configs that produced it. One lab opened its data. The other opened its interface. Neither opened both, and which one is more useful to you depends entirely on what you are trying to build.

Two disclosures, neither of them complete

MiniCPM5-2B's data release is real and it is specific. The card names the datasets, links them, and describes what each contributes — code data managed in tiers, agent SFT samples, RL samples spanning mathematics, code, general knowledge and long-context reasoning. What it does not give you is the mixture, the ordering, the decontamination procedure or the token counts per source. You can see the ingredients and not the recipe.

PAN 2's prompt release is the mirror image. The caption policy is fully specified: what a world-state description must contain, the ordering of events, how actions bind to actors, how camera motion is expressed, and a hard capacity of 1,536 text-encoder tokens with a soft target of 1,480. What it does not give you is the corpus that taught the model to read those captions — "hundreds of millions of videos annotated in the same format", with no dataset card and no release.

So if your need is to understand why a small language model behaves the way it does, MiniCPM5-2B gives you more to work with. If your need is to get a video model to produce the clip you asked for, PAN 2 gives you more. Both releases are more transparent than the median, and both stop at the point where transparency becomes competitive disadvantage.

The numbers, and what they are worth

MiniCPM5-2B is a 2,516,756,480-parameter dense Llama-architecture model — 1.98 billion parameters excluding embeddings, 42 layers, 16 query heads against 2 key-value heads, 131,072 tokens of context, English and Chinese. OpenBMB claims 2B-class open-source state of the art, with an average of 53.9 across its comparison set against 51.1 for the best larger model in that set. The individual rows are where the claim gets interesting: LiveCodeBench v6 at 69.1, AIME 2026 at 86.5, MMLU-Pro at 70.8, GPQA Diamond at 70.2, SWE-bench Verified at 46.4, τ²-Bench Telecom at 97.1, AA-LCR at 59.0 and NoLiMa at 68.1.

Those are vendor-reported on a vendor-chosen comparison set, and the set is instructive: it includes Gemma-4-E2B-it, Gemma-4-E4B-it, Qwen3.5-2B, Qwen3.5-4B, LFM2.5-2.6B and LFM2.5-8B-A1B. On AA-LCR the reported spread is 59.0 for MiniCPM5-2B against 5.3 for LFM2.5-2.6B and 17.0 for Gemma-4-E2B-it. On SWE-bench Verified it is 46.4 against 6.0, 2.0 and 3.0. Those gaps are large enough that harness differences do not explain them, but the harness is still OpenBMB's and nobody has reproduced the table independently.

PAN 2 has no table of that kind, because there is no shared instrument for video world models the way there is for code and math. What it has instead is a structured limitations section that is blunter than any vendor benchmark: about two in three simulated plans contain at least one step that is not carried out; appearance drifts across clips because each one starts from at most a single frame plus text; text actions specify dexterous manipulation only coarsely; the model may let a simulation progress while violating physical constraints, with pasta appearing in the colander before it is drained given as the example.

• What it is — MiniCPM5-2B: a dense 2.52B text model for on-device assistants, coding agents and tool workflows. PAN 2: a 29B diffusion transformer that renders the consequences of an action as video.

• Output — MiniCPM5-2B: text. PAN 2: MP4, up to 20 seconds at 704×1248 and 24 fps.

• Context or horizon — MiniCPM5-2B: 131,072 tokens, with AA-LCR 59.0 and NoLiMa 68.1 on OpenBMB's harness. PAN 2: one starting frame plus a caption, so long interactions drift and the card says so.

• Licence — MiniCPM5-2B: Apache-2.0, English and Chinese. PAN 2: Apache-2.0 on its own weights, with Qwen3.5-9B and Wan 2.2 under their own terms.

• Hardware — MiniCPM5-2B: a single consumer GPU, and the card documents vLLM, SGLang and llama.cpp paths plus a multi-chip backend. PAN 2: eight H200s by default in the vLLM recipe; four B300s for thirteen seconds of video in about 119 seconds.

The on-device claim versus the node-scale reality

MiniCPM5-2B is built to be small in the way that matters: not just few parameters, but a documented path onto hardware you already own. The card carries quickstarts for vLLM, SGLang and llama.cpp, with a note that -c 8192 sets the context length on the llama.cpp path — a line that only appears in documentation written by people who have actually run the thing on a laptop. It also documents FlagOS, a multi-chip backend that lets the same weights run across different accelerators. That is a deployment story aimed at phones, single-board computers and cheap inference boxes.

PAN 2's deployment story is the opposite in every dimension, and the reason is arithmetic rather than sloppiness. Fifty sampler steps with classifier-free guidance on every step means a hundred forward passes through a 29B transformer per clip. The recipe parallelises guidance across two ranks and sequence length across two more, then adds VAE patch parallelism across four ranks in tile mode because decoding twenty seconds of 704×1248 is its own memory problem. You do not get to run that on a laptop, and the team has not pretended otherwise.

What is worth noting is that neither model is hosted anywhere. MiniCPM5-2B has over a million monthly downloads and no first-party API; PAN 2 has no API and, as of 11 November 2026, one social post from its lab's account and nothing behind it — no endpoint, no paper link, no pricing. If you want either, you are running it.

Where the planning half can be rented

PAN 2's architecture makes one part of it separable, and that is the part most teams should not be building themselves. The conceptual reasoner — the language model that turns an observation and a requested action into the structured caption the diffusion backbone consumes — is explicitly any instruction-following model. The report ranks four by how often a simulated plan completes end to end: Claude Opus 5.5, GPT-6 Astra at medium effort, Kimi K3, then Qwen3.8-Flash. The card also notes that any instruction-following LLM can be used, and that going beyond the training distribution tends to need hand-refined prompts.

That makes the reasoner an operational choice you will want to revisit, and revisiting it is cheaper through a router than through four procurement conversations. Qwen3.8-Flash is on OrcaRouter; a single API key with automatic failover lets you run the same caption workload across several planners, at provider list price with 0% markup passed through, and keep whichever one completes your plans most often. OrcaRouter does not host PAN 2 — there is no hosted PAN 2 to route — and it does not host MiniCPM5-2B either. The renderer and the small model stay on your hardware; the planner is the commodity.

For MiniCPM5-2B the calculus is different in a way that is worth stating. On-device models exist precisely to avoid metered calls, and a 2B model that runs on a phone is not improved by routing around it. If your reason is privacy, latency or offline operation, routing is the wrong answer and MiniCPM5-2B is the right model. If your reason is cost, compare honestly: a 2B model that scores 53.9 on its own average may cost more in engineering time to reach parity on your task than a metered call to something larger.

What each release is actually for

MiniCPM5-2B is a model with a job: be the best 2B-class open model on the benchmarks its users care about, run on device, and come with the data so the community can build on it rather than around it. It is a mature release with a documentation surface that reflects several iterations — quickstarts for three serving stacks, a tool-calling section, cookbooks, fine-tuning recipes, a multi-chip backend.

PAN 2 is a research artifact with an unusually complete claim. It argues that simulative reasoning — an agent imagining the outcomes of its candidate actions before choosing one — is a capability you can build rather than a capability you hope emerges, and it implements the argument as a two-stage system with a published pattern behind it (the Generative Latent Prediction formulation in Critique of World Models, arXiv 2507.05169). The framework support that landed on 10 November 2026 made it runnable. A two-paragraph post on the lab's own X account the same afternoon did not make it launched, and nothing else has been said since.

Neither of those is a better release than the other. They are answers to different questions, published at different stages of confidence. One lab is saying "here is a small model you can trust, and here is what went into it". The other is saying "here is a system that does something new, and here is exactly where it fails". A comparison between them is only useful if you know which of those sentences describes your project.