The 'reasoning speed control' feature once seen only in big tech's proprietary APIs has finally arrived in the local open-source ecosystem! I'm talking about the Qwen3.8-27B model released by Alibaba. The most notable point of this release is 'Flexible Thinking Control,' which allows users to freely adjust the model's depth of reasoning in a local execution environment.

Developers can now dynamically control the reasoning level with the reasoning_effort parameter, and use the preserve_thinking option to maintain the reasoning context from previous steps while a multi-turn agent loop is running. This feature elegantly solves the problem of wasting tokens and latency by regenerating the thought process from scratch at every step. It has actually proven its efficiency by recording a 61.7% performance rate on SWE-bench Pro, a benchmark for solving complex code bugs, showing a significant improvement over previous models.

{
  "model": "qwen3.8:27b",
  "reasoning_effort": "high",
  "preserve_thinking": true
}

Example of vLLM and SGLang API call configuration

You might worry that it's too heavy to run locally as a reasoning model, but it's lighter than you'd think. Thanks to its hybrid attention structure, it's quite efficient, and using Unsloth's 4-bit quantized model, it runs smoothly on personal hardware (requiring about 17–19GB VRAM), such as an RTX 4090 or a MacBook with 24GB of memory. It looks like local agent development, where you can freely control a high-performance reasoning loop on your own machine without calling expensive cloud APIs, is going to get a lot more fun.