format json disables thinking?

[nshiab]

Hi,

It looks like the { format : 'json' } option disables the { think: true } option.

For example, in the code below, the model thinks and returns its response as text. It works as expected. I reused the example from the documentation.

import ollama from "npm:ollama";
import process from "node:process";

async function main() {
    const response = await ollama.chat({
        model: "qwen3:8b",
        messages: [
            {
                role: "user",
                content: "What is 10 + 23",
            },
        ],
        stream: true,
        think: true,
    });

    let startedThinking = false;
    let finishedThinking = false;

    for await (const chunk of response) {
        if (chunk.message.thinking && !startedThinking) {
            startedThinking = true;
            process.stdout.write("Thinking:\n========\n\n");
        } else if (
            chunk.message.content && startedThinking && !finishedThinking
        ) {
            finishedThinking = true;
            process.stdout.write("\n\nResponse:\n========\n\n");
        }

        if (chunk.message.thinking) {
            process.stdout.write(chunk.message.thinking);
        } else if (chunk.message.content) {
            process.stdout.write(chunk.message.content);
        }
    }
}

main();

However, if we add format: 'json' to the options, then the model doesn't think anymore. There are no thoughts anymore.

import ollama from "npm:ollama";
import process from "node:process";

async function main() {
    const response = await ollama.chat({
        model: "qwen3:8b",
        messages: [
            {
                role: "user",
                content: "What is 10 + 23",
            },
        ],
        stream: true,
        think: true,
        format: "json",
    });

    let startedThinking = false;
    let finishedThinking = false;

    for await (const chunk of response) {
        if (chunk.message.thinking && !startedThinking) {
            startedThinking = true;
            process.stdout.write("Thinking:\n========\n\n");
        } else if (
            chunk.message.content && startedThinking && !finishedThinking
        ) {
            finishedThinking = true;
            process.stdout.write("\n\nResponse:\n========\n\n");
        }

        if (chunk.message.thinking) {
            process.stdout.write(chunk.message.thinking);
        } else if (chunk.message.content) {
            process.stdout.write(chunk.message.content);
        }
    }
}

main();

Does anyone have any idea what's going on?

Thank you!

[ZoranRavic]

Setting think: true is only there to tell ollama to split the response into 2 strings: message.thinking and message.content.

It does that by detecting <think> and </think> in the response. (some models may use different separators)

If you set think: null (or omit it entirely) you'll see that the "thinking" part is in message.content.

In json mode the response only contains json without any "thinking" tokens before it,
so there is nothing to put in the message.thinking string.

What I would suggest is first asking the model to "think" and then making a second request for a better json response.