What this is
This week, Reddit's LocalLLaMA community surfaced an embarrassing finding for the open-source AI world: Alibaba's newly released Qwen 3.8 27B model—an open-source large language model that anyone can freely download and run—ships with a "botched" default parameter. The temperature parameter, the dial that controls how random an AI's responses get (higher values mean more scattered output), defaults to 1.0. The result: the model writes an entire internal monologue before editing a single line of code. Drop the temperature to 0.7, and the verbosity disappears.
Industry view
The real signal here isn't "how to tune parameters." It exposes a maturity gap in open-source LLMs. Closed-source products like OpenAI and Anthropic—the kind of AI services that don't expose their underlying weights—have product teams behind them polishing the user experience. Open-source models rely on community distribution, and what users get is often a "half-baked" product. Supporters will emphasize that open-source's flexibility allows for later tuning; critics are sharper. They point out that any company looking to build AI capabilities on open-source models must either staff a professional team to keep tuning parameters, or keep eating the cost of bad defaults. The "open-source is cheaper" story is far more complicated than the marketing suggests.
Impact on regular people
- For enterprise IT: Don't evaluate solely on benchmark leaderboards. Ask whether the default parameters actually work in your specific business scenario.
- For working professionals: When using finished products like ChatGPT or Wenxin Yiyan, simply ask for "concise answers" when the AI gets too verbose. The underlying logic is similar to adjusting temperature.
- For the consumer market: The experience gap between "domestic large models" on the market usually stems not from raw model capability, but from the depth of product polish.