DeepSeek Chat Model got issues with reasoning_content

Questions
,
1

When I use AI Agent with the chat model deepSeek, I occasionally encounter the error “The reasoning_content in the thinking mode must be passed back to the API.”

However, when I try with OpenAI, I don’t get this issue. How should I fix it?

3

Hi @Asha_Naon

This is a known issue, and the n8n team has already added it to their backlog to be fixed in a future update.
In the meantime, you can work around it by using the following community node and disabling Thinking Mode there, since the error is caused by Thinking Mode and how reasoning_content is handled.

Community node:
n8n-nodes-deepseek-v4 (install via Community Nodes in n8n)
Link: n8n-nodes-deepseek-v4 - npm

Please note: you can only install this node on self-hosted n8n instances, since unverified community nodes are not available on n8n Cloud.

4

thanks @nguyenthieutoan it’s work!!

5

Hey Asha, how did you fix it?

When I tried using the community node, my API key stopped working and when i try the official deepseek integration, it throws the same error as it did for you.

6

image
image912×732 31.1 KB

Used the community n8n-nodes-deepseek-v4 but unable to set up credential

7

@Adam_Luka @calmlion1981 The community node uses its own credential type called “DeepSeek API” - it’s separate from the built-in DeepSeek credential. After installing n8n-nodes-deepseek-v4, go to Settings > Credentials > New, search for “DeepSeek API” (the one from the community node), and paste your API key there. Then in the node, click the credential dropdown and select that new credential. The key itself from platform.deepseek.com works fine - it just needs to be added under the correct credential type.

8

@nguyenthieutoan @Adam_Luka @calmlion1981

It does work, but I’m feeling it’s not very reliable or stable, especially since the author doesn’t use English. I’ve looked into similar community nodes and it seems like they all come from the same author but under multiple accounts, which makes me even more hesitant, even though it actually solved the content reasoning issue. It looks like we should still wait for the n8n team’s latest move on this matter.

9

Totally agree, that’s just a temporary solution. I myself am very concerned about this issue, especially since DeepSeek has released powerful new models that are worth trying. The fact that it cannot be used for AI Agents is truly an error that shouldn’t exist.

10

@Asha_Naon @Adam_Luka @calmlion1981 thanks a lot for sharing your thoughts on this, I completely understand the concern around long‑term reliability of community nodes and unclear authorship.

Because I’m also heavily relying on DeepSeek for AI Agents in production, I’ve gone ahead and built and published a new community node that focuses on being stable, transparent, and well‑documented for this exact use case (DeepSeek Chat Model with AI Agent + tools, without reasoning_content issues).

The node is:

Key points:

  • It exposes a DeepSeek Chat Model (Corrected) sub‑node that is designed to work reliably with the AI Agent node and Tool Calling.

  • It avoids the whole reasoning_content class of errors by running DeepSeek in a safe configuration for multi‑turn tool flows, so you don’t hit 400s on subsequent turns.

  • The implementation is published on npm under my real identity (I’m an n8n Verified Creator), and I plan to maintain it actively as DeepSeek and n8n evolve.

How to use it on self‑host:

  • Go to Settings → Community Nodes

  • Click Install, paste: n8n-nodes-deepseek-chat-model, and install

  • Restart your n8n container so the node is loaded across the instance (especially in queue mode)

  • In your workflow, set the AI Agent’s Chat Model to DeepSeek Chat Model (Corrected)

  • Turn Thinking Mode off in that node

  • Keep your tools as usual and run your agent

This way, you get a more predictable and production‑friendly DeepSeek integration while we all wait for an official, fully supported DeepSeek AI Agent story from the core n8n team.

If any of you test this node and run into edge cases (parallel tool calls, memory backends, etc.), feel free to ping me here and I’ll be happy to iterate on it.

12

Oh man, this thing works the best outta all the DeepSeek nodes I’ve tried. Guess I don’t need to wait for n8n to patch bugs anymore , this is all I need, zero errors.