How does the mixed reasoning model work?

Let me tell you how amazing Qwen3's hybrid reasoning model is. To be honest, I was skeptical when I first saw the concept - can a model really switch between "thoughtful" and "responsive"? But as it turns out, this may be one of the most practical innovations in the history of big models. Imagine being able to answer a simple question in the blink of an eye, and when faced with a complex math problem, deriving it step-by-step like a mathematician. That kind of flexibility is simply amazing!

Thinking mode vs. non-thinking mode

At the heart of hybrid reasoning lies two distinct but complementary modes of thinking. In Thinking mode, Qwen3 "talks to itself" like a human expert - it generates its own internal reasoning process and then reaches a final conclusion. This mode is ideal for solving math problems, writing code, or answering questions that require logical reasoning. Tests have shown that Thinking Mode improves accuracy by as much as 181 TP3T on math competition questions like AIME24, while Non-Thinking Mode maintains a reasonable level of accuracy while maintaining speed when you only need a short answer.

Interestingly, the Qwen3 team found that this "bimodal switching" was not achieved by simply reducing model complexity. In fact, they deliberately designed specific data patterns and training tasks during pre-training, so that the model learns to autonomously differentiate between problems that require careful thought and those that can be responded to quickly. It's like the division of labor between System1 and System2 of the human brain.

Flexible "think budget" mechanism

What's even better is that Qwen3 can set a "thinking budget". This is like giving a model a "brain coupon" that tells it how much computational resources it can spend on thinking. Developers have found that the optimal setting often varies from task to task - writing a Python script may require a thinking budget of 2000 tokens, whereas a simple query may only require 200 tokens - and by fine-tuning this parameter, it is possible to save 40% of computation while maintaining 90% of accuracy. 40% computational cost! This is simply a golden feature in business applications.

I've personally tested the model response in Think mode. When you add the "/think" tag, Qwen3 gives you a detailed chain of reasoning before answering: "As a first step, I need to understand the mathematical structure of the question...", which is a breathtakingly clear step-by-step process. And when it switches to non-thinking mode, it gives a description of the answer in a flash - it lacks the reasoning process, but it's so fast that it completely changes the dialog experience.

The Black Technology Behind It

The secret to this ability lies in Qwen3's two-stage training strategy. The first stage of pre-training gives the model a solid foundation of general knowledge; the second stage of intensive learning specifically optimizes the thought chain generation ability. Lab data shows that the model optimized for thought chains improved its self-consistency on complex reasoning tasks by more than 30%. What's more, the team also carefully designed the "thought interruption" technique to train the model to produce reasonable outputs at different stages of computation.

I remember a user shared an interesting case on Github: when he developed a legal counseling assistant with Qwen3, he found that by dynamically adjusting the depth of thinking, he was able to quickly answer simple queries about legal terms, as well as reason in detail about the status of complex cases. The most surprising thing is that in the case of 99%, the fleeting thinking budget setting does not affect the quality of the final answer at all - this flexibility, no wonder people love it.