PAN 2 vs Qwen3.8-27B: The Opponent Is Already Inside the Model

Most model comparisons have a clean boundary: two checkpoints, two vendors, pick one. This one does not, because Qwen runs through PAN 2's plumbing in three places at once. PAN 2 loads Qwen3.5-9B as its frozen text encoder, the component that turns a structured world-state caption into the conditioning features its 29-billion-parameter diffusion transformer consumes. The PhyGround scores in PAN 2's own technical report are judged by Qwen3.8-27B, the model this page is nominally comparing it against. And Qwen3.8-Flash is one of four conceptual reasoners the report benchmarks for writing PAN 2's captions in the first place. So the honest framing is not PAN 2 versus Qwen3.8-27B. It is a video world model from MBZUAI's Institute of Foundation Models, published quietly and never given a launch, standing on Qwen components, being graded in part by a Qwen judge, compared with a 27-billion-parameter vision-language model from Alibaba that is a complete product on its own. The interesting question is which of those relationships you should care about when you are choosing what to deploy.

What the shared DNA does and does not mean

It does not mean the two models are comparable. PAN 2 outputs video; Qwen3.8-27B outputs text. Nothing in the shared encoder changes that, and no benchmark exists that ranks them against each other. Anyone presenting a head-to-head score is presenting an impression.

What it does mean is that the comparison is unusually informative about provenance. When PAN 2 reads your caption, it reads it through a Qwen encoder that MBZUAI downloaded unmodified from its own repository, and the model card is explicit that the Qwen3.5-9B weights are used as released and have not been fine-tuned by MBZUAI. The token budget the prompt validator enforces — a soft 1,480 tokens, a hard 1,536 — is Qwen3.5-9B's budget, not PAN 2's. And when the report claims a PhyGround number, that number was produced by asking Qwen3.8-27B to score generated video on a five-point scale. The judge is the rival.

That is not a scandal. Using a strong open model as an automated grader is standard practice and disclosing it is better than most labs manage. But it does mean the evidentiary chain for PAN 2 runs through Alibaba's model family at three separate points, and if you are evaluating PAN 2 you should know which of its numbers are measurements of PAN 2 and which are measurements of PAN 2 as seen by Qwen.

Qwen3.8-27B, on its own terms

Alibaba's model is a different kind of object entirely: a 27-billion-parameter causal language model with a vision encoder, built on the Qwen3.5 architecture with a hybrid layout — sixteen repeats of three Gated DeltaNet blocks followed by one Gated Attention block, 64 layers total, a hidden dimension of 5,120 and a 248,320-token padded vocabulary. Native context is 262,144 tokens, extensible to one million. It is Apache-2.0, it went up on Hugging Face on 5 August 2026, and it has 7.4 million monthly downloads.

The vendor numbers are strong and the card is one of the more complete in the open-weights field: GPQA Diamond 89.2, HLE 30.8, LiveCodeBench v6 90.3, SWE-bench Pro 61.7, Terminal-Bench 2.1 at 73.0, OSWorld-Verified 84.3, OmniDocBench 1.5 at 91.1, CharXiv reasoning 83.7, IFBench 79.5, CoWorkBench 70.7. On MathVision it reports 90.0 without chain-of-thought and 94.6 with it. Those are Alibaba's own evaluations against Alibaba's own baselines, which is the same caveat that applies to every vendor table on this site, including PAN 2's.

• Modality — Qwen3.8-27B: text, image and video in; text out. PAN 2: an optional first frame plus a structured caption in; MP4 out.

• Size — Qwen3.8-27B: 27B dense, 64 layers, roughly 27.8 billion safetensors parameters. PAN 2: a 29B diffusion transformer at 67.37 GB in BF16, plus a frozen Qwen3.5-9B encoder at 15.87 GB and a Wan 2.2 VAE.

• Context — Qwen3.8-27B: 262,144 tokens native, up to one million. PAN 2: a single caption capped at 1,536 text-encoder tokens, and one starting frame.

• Licence — Qwen3.8-27B: Apache-2.0. PAN 2: Apache-2.0 on its own weights, with third-party components under their own terms.

• Serving — Qwen3.8-27B: Transformers, vLLM, SGLang, TokenSpeed, and a hosted Qwen Cloud endpoint Alibaba lists as coming soon. PAN 2: Diffusers, vLLM-Omni and the repository's own generate.py, on your hardware, and nothing hosted anywhere.

Why the encoder choice is the most consequential line in PAN 2's card

Freezing a 9B text encoder and training only the diffusion backbone is a defensible engineering decision, and it is the reason PAN 2's diffusion half is 29B rather than 40. It also hard-caps what PAN 2 can be told. The conceptual reasoner writes a caption; Qwen3.5-9B maps it to features; the diffusion model renders from those features. Whatever Qwen3.5-9B cannot represent, PAN 2 cannot simulate, no matter how good the caption is.

This explains a limitation the card states plainly and most readers skip: the model is sensitive to prompt wording, examples beyond the training distribution relied on hand-refined prompts, and text actions specify dexterous manipulation only coarsely. Those are not three separate problems. They are the same problem seen three ways — the interface between a language description and a fixed text encoder, with a 1,536-token budget and a frozen vocabulary on one side.

Compare that to how Qwen3.8-27B handles the same job. It is the reasoner and the encoder in one object, with a 262,144-token window, native thinking control that can be disabled per request, a reasoning_effort knob and preserve_thinking for retaining reasoning context across turns. If your task is to describe a scene and a plan in enough detail that something downstream can act on it, Qwen3.8-27B is built for that and PAN 2 is built to consume the output. They are not competitors for the same slot; one is upstream of the other.

The deployment asymmetry nobody puts in the comparison table

Qwen3.8-27B is a big model that runs like a normal big model. Twenty-seven billion parameters at BF16 is roughly 55 GB of weights, which fits on one 80 GB accelerator with room for a KV cache, and the card lists four serving stacks plus an OpenAI-compatible hosted endpoint on the way. You can have it answering requests the same afternoon you decide to use it.

PAN 2 is 86.17 GB across sixty files and the serving arithmetic is done in node-minutes. Fifty sampler steps with classifier-free guidance on every step is a hundred forward passes through a 29B transformer per clip. The vLLM recipe's headline is thirteen seconds of 704×1248 video in about 119 seconds on four B300 GPUs, with guidance parallelised across two ranks and Ulysses sequence parallelism across two more; its default configuration is eight H200s in BF16 with VAE patch parallelism across four ranks in tile mode. The repository's own inference script is simpler and hungrier: one worker per GPU, each holding a full copy of the 67.37 GB transformer, each taking every Nth prompt.

Put those next to each other and the practical conclusion is not close. A Qwen3.8-27B call is a request. A PAN 2 clip is a batch job with a GPU reservation attached. If you have a twenty-second video to produce once, PAN 2 is an afternoon of setup and three minutes of render. If you have a thousand plans to simulate, you are buying time on a node and the comparison that matters is against other video world models, not against a language model.

The one place the two genuinely meet

PAN 2's intended use is simulative reasoning: an agent imagines the outcome of each candidate action, then commits to the best one. That loop needs a planner that generates candidate actions and a simulator that renders their consequences. Qwen3.8-27B is a credible planner — strong instruction following at 79.5 on IFBench, long-horizon office work at 70.7 on CoWorkBench, agentic terminal coding at 73.0 on Terminal-Bench 2.1 — and it is already in the loop, since MBZUAI used it to judge PhyGround and Qwen3.8-Flash as one of four caption writers.

If you build that loop, the planner is the part you can buy by the token and the simulator is the part you cannot. Qwen3.8-27B is on OrcaRouter, so a single API key gets you a planner with automatic failover, routing rules for which requests go where, and provider list price with 0% markup passed through — including Alibaba's own price cuts on the day they land. OrcaRouter does not host PAN 2 and cannot route it; no one hosts PAN 2. The split is not a limitation of the router, it is the shape of the market: hosted language models are a commodity, hosted video world models do not exist yet.

There is a second, quieter reason to route rather than pin the planner. PAN 2's report ranks four reasoners by full-plan completion and the ranking is not the one most people would guess — Claude Opus 5.5 first, GPT-6 Astra at medium effort second, Kimi K3 third, Qwen3.8-Flash fourth. That order was measured on MBZUAI's prompts, with their caption policy and their validator. Yours will differ. Being able to re-run the comparison without re-plumbing your stack is worth more than any single model choice in it.

What to do with this comparison

If you came here to pick one, the honest answer is that the question is malformed. Qwen3.8-27B is a general vision-language model you can serve today on one accelerator or rent by the token next week. PAN 2 is a research artifact that renders consequences, needs a multi-GPU node, has no hosted endpoint, and as of 11 November 2026 has never been launched by its own authors — one two-paragraph post on the lab's X account is the whole of the announcement. Its Hugging Face repository dates from 7 October, its GitHub repository from 5 November, and the most substantial public thing that has happened to it is vLLM-Omni's Day-0 support merging on 10 November.

The comparison worth making is narrower. If your pipeline needs a planner, Qwen3.8-27B is a strong, well-documented, Apache-2.0 choice and the decision is a routing decision. If your pipeline needs to see what an action does before taking it, PAN 2 is one of very few open-weight models attempting it, it will run on Qwen components whether you notice or not, and the honest summary of its capability is the sentence in its own limitations section: about two in three simulated plans contain at least one step that is not carried out. That is the number to plan around, and it is the one number in this entire comparison that no encoder, judge or router can improve for you.