User Reports Extreme Verbosity and Long Thinking Times in Qwen 3.8 Next Flash
A LocalLLaMA community member reported that the Qwen 3.8 Next Flash model exhibits extreme verbosity and excessively long thinking times during software engineering tasks. On single-turn coding prompts, the model spent up to 13 minutes reasoning and produced output heavy with jargon when decision-making was required. As reasoning-focused language models become more prevalent, balancing output quality with inference speed and token verbosity remains a major challenge for local developer workflows. Excessively long reasoning cycles can significantly hinder user productivity in interactive coding environments like VSCode or pi.dev. Despite high hardware throughput of around 150 tokens per second for generation and 7,000 tokens per second for prompt processing, single coding tasks could take up to an hour on platforms like pi.dev. The user noted a reluctance to reduce the thinking level due to fears of compromising performance, as seen in other Qwen 3.8 variants.
## BACKGROUND
In large language model benchmarks, prompt processing (PP) refers to reading and prefilling input context, while token generation (TG) measures the decoding speed of the generated output. Modern reasoning models generate internal 'thinking' tokens before producing a final answer, which enhances accuracy on complex logic but increases latency.