# Unpopular opinion Qwen 3.8 is hard to understand — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/6ea0a6634a419a4e

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-08-30 17:15 ingested 2026-08-31 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1w2ncr5/unpopular_opinion_qwen_38_is_hard_to_understand/)

## retained text



```
I find both Qwen 3.8 27b and Qwen 3.8 Flash Next difficult to read.

Here's some examples of what I mean:

>\*\*Model-visible tool set per turn\*\* (assembled by the host at provider-request time): persona tool allowlist ∩ session tool surface ∩ tools not \`deny\`-classified under the active permission profile.

In the above, Qwen uses the set intersection symbol as opposed to a human readable explanation. Maybe this is because it's been trained so hard on math, science, reasoning, so it's a little understandable but unnecessarily dense in my opinion.

My thoughts are that this is the consequence of minimizing "tokens per intelligence" -- that it makes reading it as a human incredibly dense.

It also uses the word "persona" which is an odd word to use when something like "mode" or "agent" would have made more sense.

And another example:

>Consent is negotiable; enforcement is gravity.

WTF does "gravity" mean in this sentence, just say "Consent is negotiable; enforcement is not."

I know a lot of claude users have been talking about how claude 5 was a step backwards in terms of human-readability. I'm somewhat afraid Qwen is taking the same road.

Qwen 3.6 was the last easy to read Qwen model IMO.

What are your thoughts?
```



document fetched 2026-08-31 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1w2ncr5/unpopular_opinion_qwen_38_is_hard_to_understand
