Users Criticize Qwen 3.8 Models for Overly Dense and Mathematical Output Style

A Reddit user has sparked discussion by highlighting that newer Qwen models, specifically Qwen 3.8 27B and Qwen 3.8 Flash Next, generate overly dense, mathematical, and cryptic language. The user attributes this shift to heavy training on reasoning and math, which prioritizes token efficiency over human readability. This highlights a growing tension in LLM development between optimizing for "tokens per intelligence" (efficiency) and maintaining natural, human-friendly language. If models become too dense or use mathematical notation instead of plain text, they may become less accessible for general users. Specific examples of Qwen 3.8's output include using set intersection symbols (∩) to describe permission logic and using cryptic metaphors like "enforcement is gravity" instead of straightforward phrasing. The user noted that this trend of declining readability was also observed in other advanced models like Claude 5, whereas Qwen 3.6 was much easier to read.