gemma4:e4b beats qwen3.5:9b on speed, but both ignore summary format

This post is the sixth in a series that had previously covered only coding-related tasks, so it shifts the lens to drafting and summarization. The method was deliberately simple: the author fed a roughly 8,400-character technical article of their own into Ollama's API with no background load and asked for a 3–5 bullet Japanese summary.

The most striking result is not speed but the empty response from qwen3.5:9b. Its done_reason of "length" indicates the model consumed the entire generation cap on thinking, a failure the author says reproduced with the Ornith-family models in study4 and study5. Setting think: false fixed it, making the case that thinking-capable models used directly through an API carry a risk of blank output unless that flag is explicit.

On quality, both models were accurate in content but ignored the format instruction, producing long strategic reports with emoji headings, comparison tables, and even code blocks. The author's practical lesson is that 'be concise' alone does not work; prompts need hard limits on character count and output format. They also flag that the test was a single run on one task and that different document types or prompts could change the instruction-following results.