Articles

In the PlyLoom arena: Qwen versus DeepSeek

Two models got the same question, and their answers were opened as PlyLoom maps. What a map shows at once and a text answer hides — a retelling of a small experiment the author of PlyLoom described on Habr (in Russian).

· Matvey Zakharenko · 4 min read

How it was set up

Two models — Qwen 3.8 Max and DeepSeek V4 — got the same cross question: tell about yourself and about your rival. The prompt was the one PlyLoom prepares for a new map from materials, and its materials block was left as an empty placeholder. The answers were opened in PlyLoom and compared by what shows on the map: which sheets and nodes came out, whether there were twins and open questions. Both answers were also exported as Mermaid diagrams and viewed in mermaid.live.

What the two maps held: the Qwen answer and the DeepSeek answer Qwen 3.8 Max Released models Key technologies, briefly Singled out: low training cost One kind of metric Duplicates: the same entity made again on another sheet Little about DeepSeek A short answer: a few simple entities. DeepSeek V4 Many open questions and caveats Two risks One twin: “Qwen (mentioned in the request)” The materials block is empty — no filling from memory “Qwen is the closest rival” — a hypothesis Suggests three ways to work A wordy, cautious answer. Strictly from sources A draft from general knowledge Mixed PlyLoom 0.14:Where facts come from
What the two maps held. The diagram is ours, from the description in the original article; the exact nodes are only in its screenshots.

What Qwen answered

Briefly. The model listed its released models, described the key technologies in short and singled out the low cost of training. There were few entities, all simple, and one kind of metric. About DeepSeek it said little, where it could have said more, and in a more structured way.

The map shows at once what is easy to miss in a text answer: Qwen made the same entity again on different sheets instead of putting one node on two sheets. In PlyLoom these are different cards, and an edit to one does not reach the other — such duplicates have to be merged by hand.

What DeepSeek answered

At length and with doubts: many open questions, stated uncertainty and two risks. Exactly one twin — the node “Qwen (mentioned in the request)”. At first it looked as if DeepSeek knew almost nothing about Qwen, and the author asked it why.

The answer was more interesting than the map. The model had noticed that the materials block in the prompt was empty and would not fill the map from memory: PlyLoom's rules treat materials as sources, facts and dates are not made up, and the unknown is marked as unknown, a hypothesis or “source needed”. It explained that it had not claimed not to know Qwen — only that the materials said nothing about it — and it filed “Qwen is DeepSeek's closest rival” under hypotheses. Then it suggested three ways to work: strictly from sources, a draft from general knowledge marked “unverified”, and mixed — and asked which sides of Qwen mattered: versions, licences, benchmarks, API, prices, open weights, modalities, the ecosystem.

What follows

Roughly a draw. Qwen answered the question, but sparingly and with duplicates. DeepSeek decided it had been restricted, and said so honestly. The main point is not the models but the format: a map shows the difference in a minute — by the sheets, the twins and the number of open questions. In a text the same things would have to be dug out.

For PlyLoom the conclusion is practical. If models read an empty materials block differently, the choice has to be explicit. In 0.14 the AI → A new map… window has a step for where the facts come from: only your materials, a draft from general knowledge (marked as unverified), or mixed. The rule about sources goes into the prompt before the task, and the model no longer has to guess what is expected.

A caveat

This is a small informal experiment, not a benchmark: one question, one answer from each model, judged by eye. We add no numbers that are not in the original article.

How to repeat it

  1. AI → A new map… → the Explore a topic task → choose where the facts come from → Copy the prompt.
  2. Paste the prompt into two different AI chats and ask the same question.
  3. Paste each answer into the same window — PlyLoom shows a preview of the map. Keep both maps and compare them: sheets, twins, open questions. For an article, use Project → Assemble a document…, the Mermaid diagram tab.
Open PlyLoomThe original on Habr (RU)

The original article, with screenshots of both answers and the discussion, is on Habr (in Russian), 12 October 2026.