qwen3 support

[luxixing]

Validations

• I believe this is a way to improve. I'll try to join the Continue Discord for questions
• I'm not able to find an open issue that requests the same enhancement

Problem

this is qwen3 release blog：https://qwenlm.github.io/zh/blog/qwen3/
qwen3 is best for mcp or tools use
i hope continue support qwen3

Solution

No response

[ccckblaze]

wait for update

[TyDunn]

From Discord:

there seems to be some issues with Qwen3 out of the box, looking at the github issues. for one, the reasoning is not displayed at all. I activated the reasoning parser in vllm, which means that the reasoning content is in the completion.choices[0].message.reasoning_content of the response? IIRC these kind of "features" in the UI were handled by checking the model name often (e.g. looking for "anthropic", "r1" or "o1" in the model name?) I couldn't find a setting in the config yaml to explicitly mark it as a reasoning model. maybe there's an easy hotfix?

[TyDunn]

It looks like it works with both thinking and no_think via Ollama: https://hub.continue.dev/ollama/qwen3-14b

What model runners are you using? The one on Discord was using vLLM

[antony66]

There certainly is an issue with qwen3 support. I am using ollama and qwen3:32b model. I tried ollama on mac os (m4 with 128 Gb unified ram) and linux (2 x A100 80 Gb). In both cases reasoning just stops after a certain time, like 2.5 minutes. No errors on ollama side (OLLAMA_DEBUG=1). The same prompt is working fine via Page Assist browser plugin. Moreover, the same prompt via CLINE is also working fine. Which makes me think this is continue.dev issue.

[metcalfc]

I've confirmed that Qwen3 works with Ollama and Continue. I tried the 4B and 30B models. Thinking and no_think work in Continue.

name: Qwen3 via Ollama
version: 1.0.0
schema: v1
models:
  - uses: ollama/qwen3-30b
    override:
      roles:
        - edit
        - chat
        - apply
        - autocomplete
  - uses: ollama/nomic-embed-text-latest
context:
  - uses: continuedev/diff-context
  - uses: continuedev/codebase-context
  - uses: continuedev/url-context
  - uses: continuedev/folder-context
  - uses: continuedev/terminal-context
  - uses: continuedev/code-context
  - uses: continuedev/file-context

[NGC13009]

At present, the qwen3 running on ollama will be prompted that it cannot set the autocomplete promptTemplates, but in reality, it seems to be possible to set them, and the continue plugin has not reported any errors. I think this may be a bug in the configuration file rules.

In addition, currently continue still does not support the switch to set the thinking mode for qwen3 running on ollama. I can only manually turn on or off thinking using the /notlink prompt.

Hope to further increase support, including:

1. When chatting, it is more convenient to turn on or off the QWEN3 series of thinking (or is there a better way? It can support a configuration option for the model. When it is true, the thinking mode switch is forcibly displayed to be compatible with any possible model name)
2. Autocomplete can be configured to turn off thinking mode (I am currently using a custom prompt implementation, but I hope there is a better implementation effect), and it is also compatible with automatically ignoring the '' tag when turning off thinking mode.
3. Increase compatibility with the Alibaba Bailian platform API. Currently, continue is still not compatible with the Alibaba API's thinking model to display the thinking chain function, which appears to be stuck during loading and does not display the thinking process. Tracking: # 5037

At present, qwen3:4b is a model that is very suitable for running on small graphics cards (<8GB), and after experiencing it, it is very suitable as an autocomplete code model and chat model. I hope to optimize the experience for qwen3 as soon as possible, which will be very helpful : )

[antony66]

How do you guys set it to no_think mode?

P.S. found the answer myself: just add "/no_think" to the prompt anywhere

[joakimriedel]

2. Autocomplete can be configured to turn off thinking mode (I am currently using a custom prompt implementation, but I hope there is a better implementation effect), and it is also compatible with automatically ignoring the '' tag when turning off thinking mode.

I've been experimenting with this for qwen3-30b-a3b hosted in LM Studio using a custom autocomplete template to inject the /no_think in the prompt but I have not yet succeeded. Your image do not fully convey how to do this properly in the yaml. I want to be able to have thinking in the chat/edit modes, but no thinking in the autocomplete. It would be nice with first class support for this in Continue.

[NGC13009]

@joakimriedel

  - name: qwen3:4b-16k - ollama
    provider: ollama
    model: qwen3:4b-16k # usermodel: 16k ctx size
    apiBase: http://127.0.0.1:11434/
    roles:
      - chat
      - edit
      - autocomplete
      - apply
      - summarize
    promptTemplates:
      autocomplete: |
        `
        按照给出的文本前后缀，直接给出猜测的补全片段。<|fim_prefix|>{{{prefix}}}<|fim_suffix|>{{{suffix}}}<|fim_middle|>当前文件名称：{{{filename}}}, 项目名称：{{{reponame}}}/nothink
        `  # chinese prompt meaning: Provide a guessing completion fragment directly based on the given text prefix and suffix. .... current file name ....... project name ........
    requestOptions:
      verifySsl: false

[hr0109]

@joakimriedel

• name: qwen3:4b-16k - ollama
  provider: ollama
  model: qwen3:4b-16k # usermodel: 16k ctx size
  apiBase: http://127.0.0.1:11434/
  roles:
  • chat
  • edit
  • autocomplete
  • apply
  • summarize
    promptTemplates:
    autocomplete: |
    按照给出的文本前后缀，直接给出猜测的补全片段。<|fim_prefix|>{{{prefix}}}<|fim_suffix|>{{{suffix}}}<|fim_middle|>当前文件名称：{{{filename}}}, 项目名称：{{{reponame}}}/nothink # chinese prompt meaning: Provide a guessing completion fragment directly based on the given text prefix and suffix. .... current file name ....... project name ........
    requestOptions:
    verifySsl: false

@NGC13009 我用ollama官方的qwen3:4b模型+您贴出来配置文件+vscode商店的最新版continue。可以正常聊天，但是自动补全结果永远是</think>。您可以提供您用的模型出来给我们参考下吗？感谢！