Qwen 3.8 27B 非常出色，但默认会过度思考到离谱的程度
16th August 2026
2026年8月16日
Friday’s big release was Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba’s Qwen research lab. I’ve been looking forward to this one: 27B is an excellent size for running a model on a reasonably specced laptop, and its predecessor Qwen 3.6 27B was impressive.
周五的重磅发布是 Qwen 3.8 27B，这是阿里巴巴 Qwen 研究实验室推出的、采用 Apache 2 许可的 27B 参数、具备视觉能力的 LLM。我一直很期待这个模型：27B 是在配置合理的笔记本电脑上运行模型的绝佳规模，而它的前代 Qwen 3.6 27B 令人印象深刻。
Qwen’s self-reported benchmarks for this model are eye-opening. They show a boost from both Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, which was one of Qwen’s strongest models of any size as recently as May this year. It will be interesting to hear what independent benchmarks have to say about the model.
Qwen 针对该模型自行报告的基准测试结果令人大开眼界。它们显示，相比 Qwen 3.6 27B 以及非开放权重的 Qwen 3.7-Plus 都有提升；后者直到今年 5 月都还是 Qwen 各种规模中最强的模型之一。听听独立基准测试对这款模型有何评价，会很有意思。
I’ve been running the model on two different machines: my 128GB M5 Max MacBook Pro, and an NVIDIA DGX Spark. On both machines I’m running LM Studio and their 17GB Q4_K_M quantized build. I also tried using llama-server directly on the Spark.
我一直在两台不同的机器上运行这个模型：我的 128GB M5 Max MacBook Pro，以及一台 NVIDIA DGX Spark。两台机器上我都运行 LM Studio，并使用 其 17GB Q4_K_M 量化版本。我还尝试在 Spark 上直接使用 llama-server。
The default of extra high results in spectacular over-thinking
默认的 extra high 会导致极其夸张的过度思考
Qwen’s documentation describes the model as defaulting to xhigh for the reasoning effort, and the LM Studio GGUF I’ve been trying preserves that default:
Qwen 的文档说明，该模型的推理强度默认设为 xhigh，而我一直在试用的 LM Studio GGUF 保留了这个默认值：
Qwen3.8 comes with official support for
reasoning_effort, which can be used to adjust reasoning depth and control cost:Qwen3.8 官方支持
reasoning_effort，可用于调整推理深度并控制成本：
xhigh(default): for complex tasks demanding thorough analysisxhigh（默认）：用于需要透彻分析的复杂任务medium: balancing accuracy and speedmedium：平衡准确性与速度low: efficient reasoning optimizing for speed and costlow：高效推理，优化速度与成本
This is a hilarious default. It’s absolutely not a good way to run t...