Qwen samplers spark min_p debate

The interesting bit is not whether min_p is “better,” but that sampler defaults are becoming part of the model contract.

• –Qwen’s official generation config lists temperature 1.0, top_k 20, and top_p 0.95, with no min_p field, so replacing them means leaving the tested path.
• –Reasoning models may need broader token diversity during hidden or explicit thinking, while aggressive truncation can make them loop, collapse, or over-prune useful low-probability branches.
• –min_p is attractive for local users because it adapts to confidence, but support is uneven across runtimes and the strongest evidence for it predates today’s reasoning-heavy models.
• –For developers serving local LLMs, this is a reminder to benchmark sampler changes against task quality, not just vibe-test chat outputs.