PAN 2 對比 Gemma 4 12B：模擬器與推理器並非競爭對手

Every comparison page on the internet wants a winner, and this one cannot honestly have one, because PAN 2 and G​emma 4 12B do not perform the same task. PAN 2 is a 29-billion-parameter video diffusion transformer from MBZUAI's Institute of Foundation Models that takes an image and a description of actions and returns up to twenty seconds of 704×1248 video showing what happens next. G​emma 4 12B Unified is an 11.95-billion-parameter encoder-free multimodal language model from G​oogle DeepMind that takes text, images and audio and returns text. One renders consequences; the other talks about them. If you arrived looking for a benchmark table that ranks them against each other, none exists and none could — they share no evaluation. What does exist is a real decision, because the two are increasingly deployed in the same pipeline: the reasoner writes the plan, the simulator shows you what the plan does. So the useful question is not which is better but which one your problem actually needs, and what each costs you to hold.

直白陳述的不一致

G​emma 4 12B has a public evaluation table and PAN 2 does not, and that asymmetry is not a quality signal — it is a category signal. G​oogle publishes MMLU Pro, GPQA Diamond, AIME, LiveCodeBench and a dozen other text and vision benchmarks because language models are measured on a shared yardstick that has existed for years. PAN 2's card publishes Physics-IQ Verified, PhyGround, RBench and a human preference study because video world models are measured on instruments that were invented recently, mostly by the labs building the models.

• Parameters — PAN 2: 29B in a single multi-modal diffusion transformer, with a frozen Qwen3.5-9B text encoder and a Wan 2.2 VAE loaded alongside. G​emma 4 12B: 11.95B, 48 layers, encoder-free, with raw image patches and audio waveforms projected straight into the decoder's embedding space.

• Inputs and outputs — PAN 2: an optional first frame plus a structured world-state caption, out comes an MP4. G​emma 4 12B: text, image, audio and video in, text out, with a 256K-token context and native system-prompt support.

• Licence — PAN 2: Apache-2.0 on its own code and weights, with third-party components under their own terms. G​emma 4 12B: Apache-2.0 under the G​emma 4 licence terms.

• Weights on disk — PAN 2: 86.17 GB across sixty files in the IFM/PAN2 repository, of which 67.37 GB is the transformer itself and 15.87 GB the text encoder. G​emma 4 12B: about 12 billion parameters of BF16, roughly a quarter of the transformer alone.

根據其自身證據，每一個實際上擅長什麼

G​emma 4 12B's numbers are the more legible of the two and they are respectable for the size. Vendor-reported MMLU Pro of 77.2%, GPQA Diamond of 78.8%, AIME 2026 without tools at 77.5%, LiveCodeBench v6 at 72.0% and a Codeforces rating of 1659 put it well clear of the previous generation — G​emma 3 27B scores 67.6%, 42.4%, 20.8%, 29.1% and 110 on the same rows. On vision it posts MMMU Pro 69.1% and MATH-Vision 79.7%. On long context, MRCR v2 with eight needles at 128K lands at 43.4%. Those are G​oogle's own evaluations of G​oogle's own model, and the card says so; treat them as a well-documented vendor claim rather than an independent verdict.

PAN 2 has nothing comparable and does not pretend to. Its claim is narrower and stranger: that a diffusion model conditioned on a structured, time-grounded description will simulate the physical and causal consequences of an action well enough for an agent to compare candidate plans before committing to one. The support for that claim is a 300-case single-action study across six capability categories, a 200-prompt multi-step study with 456 steps, and a 1,400-comparison blind human preference test. All three are the team's own. None of them transfer to anything G​emma 4 12B is measured on.

部署問題正是這項比較真正見真章的地方

在這裏，這兩個模型不再無從比較，因為兩者都得跑在你付錢買的硬體上。

G​emma 4 12B runs on a single consumer or workstation GPU at BF16 and on considerably less with quantization. It has day-one support in the standard inference stacks, a 262K vocabulary, and a documented set of sampling and thinking-mode defaults. If your workload is document understanding, tool calling or reasoning over a long context, this is a boring, well-trodden deployment and the interesting decisions are all about prompting.

PAN 2 的部署一點也不無聊。vLLM 配方本身的標題就是：十三秒的 704×1248 影片，在四張 B300 GPU 上大約需要 119 秒，並將無分類器引導分散在兩個 rank 上平行處理，Ulysses 序列平行則分散在另外兩個 rank 上。預設配置是八張 H200，使用 BF16。取樣器執行五十步，且每一步都套用引導——每個片段要對 29B transformer 做一百次前向傳遞——這就是為什麼平行化不是可有可無的選項。而且條件格式不是自由文字：模型期望的是結構化的世界狀態描述，而儲存庫附帶一個提示重寫技能，內含驗證器，會對照文字編碼器的預算，強制執行 1,480 個 token 的軟性上限與 1,536 的硬性容量。

That last detail is the one that decides most evaluations. You cannot type a wish into PAN 2 and get a video. You have to describe events in order, bind actions to actors, and specify camera motion, and the model card is explicit that the agent case studies relied on hand-refined prompts. G​emma 4 12B will happily take a paragraph of prose and give you something useful back on the first try. If you want to know how PAN 2 feels in practice, the honest answer is that it feels like a model with an interface contract, not a chat box.

兩者在同一系統中的歸屬位置

The reason this comparison keeps coming up is that agent work now needs both halves. An agent that plans in text and acts in a real environment needs a reasoner to decide what to do and, increasingly, a simulator to preview what doing it would look like. G​emma 4 12B can be the reasoner — it is small enough to co-locate with the rest of a stack, it handles images and audio as inputs, and its instruction-following and tool-use scores are the point of the model. PAN 2 can be the simulator, and its own architecture assumes something like G​emma is already in the loop: the conceptual reasoner is defined as any instruction-following language model, and the report ranks four of them by how often a simulated plan completes end to end — Claude Opus 5.5 first, then GPT-6 Astra at medium effort, then Kimi K3, then Qwen3.8-Flash.

If that is the shape of your system, the pairing is not PAN 2 versus G​emma 4 12B at all. It is PAN 2 plus a reasoner, and the choice of reasoner is where the cost and the quality variance actually live.

這正是路由器真正發揮作用、而非只做行銷噱頭的地方。模擬的那一半必須跑在你自己的 GPU 上——沒有人代管 PAN 2，連 MBZUAI 也不例外——但推理的那一半可以是按用量計費的 API 呼叫，而在 Claude Opus 5.5、GPT-6 Astra 與 Qwen3.8-Flash 之間切換，以找出哪一個最常完成你的計畫，這是路由問題，不是採購問題。Qwen3.8-Flash 已在 OrcaRouter 上，因此你可以透過一組 API 金鑰，搭配自動容錯移轉，以供應商定價，將三種規劃器對上你自己的 PAN 2 部署進行基準測試，並享有0% 加價轉嫁。OrcaRouter 不代管 PAN 2，也無法路由它。它可以讓管線中便宜的那一半保持彈性，而昂貴的那一半則固定綁在硬體上。

什麼會改變這個比較？

有兩件事，而截至 2026 年 11 月 11 日，這兩件事都不存在。第一，由非 PAN 2 作者的人對 PAN 2 進行獨立評估——在那之前，其能力聲明屬於廠商自行報告，也應如此解讀。第二，PAN 2 的代管端點或正式發行說明——該實驗室 X 帳號上的那則貼文不算——這會把自架決策變成價格比較，並讓本頁的部署章節在一夜之間過時。

G​emma 4 12B needs neither. It is a shipped, documented, widely deployed model with over 1.8 million monthly downloads on Hugging Face, and its evaluation table has had months of community use to find the soft spots. The risk profile of adopting it is close to zero. The risk profile of adopting PAN 2 is that you spend a week on infrastructure for a model whose limitations section says about two in three simulated plans contain at least one step that is not carried out. That sentence is not a reason to skip it. It is a reason to know what you are buying before you rent the GPUs.

如何選擇

If you need text, images or audio understood and answered, PAN 2 is not a candidate and no amount of GPU budget makes it one. Take G​emma 4 12B, or one of the larger G​emma 4 variants if the 12B's numbers are not enough. If you need to see what an action would look like before you take it — robotic manipulation, egocentric task simulation, physical-reasoning evaluation — G​emma 4 12B cannot do that either, and PAN 2 is one of the few open-weight models that tries. If you need an agent that does both, the answer is a stack, not a model, and the interesting engineering is in the interface between them: the structured caption format that PAN 2 demands is exactly the artifact a reasoner like G​emma 4 12B would have to be taught to produce.