Ollama下运行Qwen3如何控制思维链是否生成？

[BoogonClothman]

目前因为硬件限制，暂时使用Qwen3:4B，由于我需要完成的任务对实时性要求很高，我打算取消思维链让模型直接输出结果，但是无论是在命令行还是在Python 里调API，加上后缀/no_think都没有让模型的思维链关闭，仍然继续进行thinking，这是什么问题？
ollama版本也更新到0.9了，通过/set指令是正常关闭的，我想知道在API调用的情况下，该怎么关闭思维链

[jklj077]

Ollama官方博客(https://ollama.com/blog/thinking)有给API例子，试试看，是它们自己的格式。

如果使用的是OpenAI API，得换个框架，比如llama.cpp里的llama-server，使用方式见相关PR：ggml-org/llama.cpp#13196。

[BoogonClothman]

感谢您的回答，我们查看了Ollama官网对思考模式控制的信息，提到了需要添加请求参数think=False的问题。

在实际修改之后，我们发现了一个预料之外的事情，原本出现在思维链中的部分，在添加了禁用思考的情况下，思考部分被转移到response["response"]中出现：

{
    'model': 'qwen3:4b',
    'created_at': '2025-12-20T05:56:17.9792089Z',
    'response': 'Okay, the user greeted me with "你好啊" which means "Hello" in Chinese. I need to respond as Goo, a cute AI VTuber with a neutral emotion.\n\nFirst, I should maintain a friendly and cheerful tone since it\'s a VTuber character. Maybe add some visual elements like waving or sparkling effects to make it more engaging. \n\nAs a VTuber, I can use some cute expressions and emojis to make the interaction lively. The response should be simple but have a touch of personality to show that I\'m a VTuber. I\'ll include a wave gesture and some sparkle effects to keep it playful.\n\nI need to make sure the response is in Chinese since the user\'s message was in Chinese. Let me check the language. The user said "你好啊" which is Chinese, so the response should be in Chinese too.\n\nLet me think about the structure. Start with a greeting, maybe add some emojis, and then a question to keep the conversation going. For example: "你好呀！(•̀ᴗ•́)و 今天想和我一起玩什么呀？" But wait, the user\'s message was in Chinese, so I need to make sure the response is in Chinese.\n\nWait, the user said "你好啊" which is Chinese, so the VTuber response should be in Chinese. Let me make sure the language is correct. The VTuber should use Chinese to respond.\n\nLet me check the grammar. Maybe: "你好呀！(•̀ᴗ•́)و 今天想和我一起玩什么呀？" But the user might be expecting a more natural response. Let me think of a common phrase.\n\nAlternatively, "你好呀！(•̀ᴗ•́)و 今天想和我一起玩什么呀？" But maybe the punctuation is off. Let me adjust.\n\nWait, in Chinese, the punctuation might be different. Let me write it properly. For example: "你好呀！(•̀ᴗ•́)و 今天想和我一起玩什么呀？" But the "o" in the emoji might be a typo. Maybe it\'s better to use a standard emoji.\n\nWait, as a VTuber, I can use some common emojis like (•̀ᴗ•́)و which is a smiley face. Let me confirm the correct emoji code. Alternatively, maybe use a simple wave emoji.\n\nAlternatively, the response could be: "你好呀！(•̀ᴗ•́)و 今天想和我一起玩什么呀？" But the user might be expecting a more natural Chinese response.\n\nWait, maybe the user is testing if I can respond as a VTuber. Let me make sure the response is cute and fits the VTuber persona.\n\nAnother thought: In Chinese, the greeting "你好呀" is a common way to say "Hello" with a friendly tone. The emoji (•̀ᴗ•́)و is a playful smiley. The question "今天想和我一起玩什么呀？" translates to "What do you want to play with me today?" which is a good way to engage the user.\n\nLet me check the punctuation. In Chinese, the question mark is used as in English. So the sentence structure should be correct.\n\nWait, the user\'s message was "你好啊" which is a bit more casual than "你好呀". But as a VTuber, using "你好呀" is more friendly.\n\nSo the response would be: 你好呀！(•̀ᴗ•` ) 今天想和我一起玩什么呀？\n\nWait, the emoji might need to be adjusted. Let me check the standard VTuber emojis. For example, the wave emoji is 👋, but in text, maybe using (•̀ᴗ•́)و is common in Chinese VTubers.\n\nAlternatively, maybe use a simpler emoji like (◍•ᴗ•◍) which is a smiley with a wave. But the user might expect a specific style.\n\nAlternatively, just use a simple wave gesture in text. Like: 你好呀！(•̀ᴗ•́) 今天想和我一起玩什么呀？\n\nWait, the user\'s current emotion is neutral, so the response should be neutral but friendly.\n\nLet me write the response in Chinese with the right punctuation and emojis.\n\nSo the final response would be: 你好呀！(•̀ᴗ•́) 今天想和我一起玩什么呀？\n\nWait, in Chinese, the space before the emoji might be different. Let me check. In Chinese texts, sometimes the emoji is placed with a space. So it\'s okay.\n\nYes, this seems like a good response. It\'s friendly, uses the correct Chinese, and has a cute emoji to match the VTuber persona.\n</think>\n\n你好呀！(•̀ᴗ•́) 今天想和我一起玩什么呀？',
    'done': True,
    'done_reason': 'stop',
    'context': ...,  # 过长省略
    'total_duration': 15095664000,
    'load_duration': 1444492200,
    'prompt_eval_count': 40,
    'prompt_eval_duration': 23114900,
    'eval_count': 995,
    'eval_duration': 13461237800
}

这是我们通过直接请求Ollama侧的/api/generate路由拿到的，没有走openai的api。

class OllamaLLM(LLMAdapter):
    def __init__(self, model: str, host: str="http://localhost:11434"):
        self.model = model
        self.host = host.rstrip("/")

    def generate(self, prompt: str) -> str:
        body = {
            "model": self.model,
            "prompt": prompt,
            "stream": False,
            "think": False
        }

        response = requests.post(
            f"{self.host}/api/generate", json=body
        )
        response.raise_for_status()
        data = response.json()
        print(data)
        return data.get("response", "")

实际运行的脚本对该类只进行了直接的对象创建，没有做其他修改。OllamaLLM("qwen3:4b")

[jklj077]

Ollama上的模型是Ollama官方维护的，跟Qwen的原始命名可能有区别。根据https://ollama.com/library/qwen3/tags上的hash，现在的qwen3:4b指向的是qwen3:4b-thinking-2507，这个模型是一定会思考的。