Qwen 3.8 Max and DeepSeek V4 Compared on a Single Cross-Domain Question in PlyLoom

An author on Habr ran a small experiment. They took two models, Qwen 3.8 Max and DeepSeek V4, and put the same cross-domain question to both. They then loaded the graph-format answers into PlyLoom so they could compare structures rather than paragraphs of text. Two models, one query, one analysis tool. The announcement is short: it does not include the question itself, any quantitative results or any conclusions, which remain in the full version of the article.

What we know about the experiment

The announcement tells us little, but what it does say is clear. The model version names are given as the author stated them; we did not verify them independently. A "cross-domain" question, judging by the wording, requires connecting several fields or concepts. Which ones is not stated. A graph answer is a set of nodes and the links between them, not continuous text. In this setup, PlyLoom serves as the platform where the results from both models are loaded for analysis and comparison.

• Models: Qwen 3.8 Max and DeepSeek V4 (names as given in the source).
• Query: the same cross-domain question for both.
• Output format: graph answers.
• Comparison tool: PlyLoom.
• Publication: Habr, 12 October 2026.
• Not disclosed: the text of the question, the number of nodes and links, the evaluation criteria, the number of runs, the final conclusion.

How graph comparison differs from the usual kind

Comparing the text answers of two models is awkward. Length, style and a confident tone get in the way: a persuasively written paragraph is easily mistaken for a more accurate one. A graph strips out some of that noise. It shows which concepts the model picked out, which links it drew between them and what it left out. Two models can be compared by the set of nodes, by link density and by where their pictures of the world diverge.

The approach also has a weak spot. A graph is itself a product of generation. Its shape depends on how the query is worded, on the required output format and on how the answer is parsed into nodes and edges. If one model was asked to produce the structure directly and the other's answer was parsed after the fact, the comparison will be skewed. And one question yields one data point. It shows how the models behaved on that question; it does not become a ranking.

How to read such comparisons and repeat them yourself

Before drawing conclusions from an experiment like this, check a few things. Were the conditions identical: the prompt text, system instructions, temperature, whether reasoning mode was on or off? How many runs were there? A single model answer depends heavily on chance. Who turned the answer into a graph, and how: by hand, with a script or with the model itself? Finally, were the facts inside the nodes checked? A neat link between two concepts stays neat even if it is made up.

For those who want to repeat the experiment, a simple scheme will do. Take a question from your own field, where you can judge correctness yourself. Fix the run parameters and save the raw answers before any processing. Run the query several times and add two or three more questions of a different type. Only then compare the graphs. What the Qwen and DeepSeek experiment actually showed can be found in the full article on Habr. For now, the public announcement contains no data that would let anyone name a winner.