Qwen3.5 user pushes back on overthinking

My read: the "overthinking" reputation looks at least partly like a reproducibility problem, not a model flaw. The same Qwen3.5 can look crisp or chaotic depending on the sampler, prompt template, and tool stack around it.

• –The poster keeps the setup intentionally small: defaults only, one local server, and four simple tools.
• –That makes the anecdote useful because it strips out a lot of agentic complexity that can inflate reasoning traces.
• –Qwen's own model card says Qwen3.5 thinks by default and recommends specific sampling settings, so "pure defaults" is a legitimate baseline.
• –The post is a reminder that local-model anecdotes are hard to compare unless people share prompts, temperatures, top-p values, context size, and tool definitions.
• –This does not disprove the overthinking complaints, but it does suggest they may be workflow-specific rather than universal.