DeepSeek chat and reasoning quick start with four examples

DeepSeek chat and reasoning quick start with four examples

One canvas with four ways to reach DeepSeek from n8n. A local model on Ollama answers the chat. Beside it sit a call to
DeepSeek's own API with thinking on, the same call with thinking off, and a conversational agent with memory. Only the
local model is joined to the chat. The other three are there to copy from or to join up yourself.

What is inside

• workflow.json: the workflow as n8n exports it, fifteen nodes. Seven are sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The links to four credentials in its author's n8n are gone, so select your own.
• The two API calls and the agent's model asked for deepseek-chat and deepseek-reasoner, which DeepSeek announced it
  would discontinue on 24 July 2026. All three now ask for deepseek-flash. The thinking-on call sends
  "thinking": {"type": "enabled"} and the thinking-off call sends "thinking": {"type": "disabled"}, which is how
  DeepSeek's change log says those two names behaved.
• In DeepSeek JSON Body the system message held your chat text and the user message said "Hello!". They are the right
  way round now: the system message says "You are a helpful assistant." and the user message is your chat text.
• Both HTTP bodies pasted your chat text between quotation marks, so a quote or a line break in it broke the JSON. They
  now use JSON.stringify on it.
• The sticky notes were reworded to say which example is joined to the chat, and the long one about the API now lists
  the current model names.
• The wiring is as the original copy had it: only the Ollama chain is joined to the chat trigger.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• For the local example: an Ollama server that n8n can reach, with the model pulled: ollama pull deepseek-r1:14b. It
  is a 9.0 GB download, and Ollama DeepSeek asks for a context of 16,384 tokens.
• For the other three: a DeepSeek API key from https://platform.deepseek.com/api_keys.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. The local example: add an Ollama credential (Base URL http://localhost:11434 unless you changed OLLAMA_HOST) and
   select it in Ollama DeepSeek. Save, open the chat on the canvas and write a message.
3. The two HTTP examples: add a Header Auth credential. Its name is Authorization and its value is Bearer followed
   by your key. Select it in DeepSeek Raw Body and DeepSeek JSON Body.
4. The agent: add an OpenAI credential whose Base URL is https://api.deepseek.com and whose key is your DeepSeek key.
   Select it in DeepSeek.
5. To try one of those three, draw a line from When chat message received to its first node. Remove the line to the
   Ollama chain first, so that only one example answers.

How a request runs

1. When chat message received passes your message to Basic LLM Chain2 as chatInput.
2. The chain sends it to Ollama DeepSeek, which runs deepseek-r1:14b with a temperature of 0.6. The system message
   is "You are a helpful assistant."
3. The reply goes back to the chat.

The three unjoined examples work the same way once you join them. DeepSeek Raw Body and DeepSeek JSON Body post to
https://api.deepseek.com/chat/completions. The agent keeps the last five exchanges in Window Buffer Memory.

Change the result

• The local model: Ollama DeepSeek, with any model you have pulled. Its options hold the context size and the
  temperature.
• The API model: the model line in each HTTP body, or the model in DeepSeek. deepseek-v4-pro is DeepSeek's other
  current model.
• Thinking: "type": "enabled" or "disabled" in the HTTP bodies. Thinking is on by default at DeepSeek, and a model
  in thinking mode ignores temperature.
• The assistant's instructions: the system message in Basic LLM Chain2, AI Agent or DeepSeek JSON Body.

Good to know

• Model names: DeepSeek's change log, read on 12 November 2026, names deepseek-flash and deepseek-v4-pro as the
  current models. It lists deepseek-chat and deepseek-reasoner as due for discontinuation on 24 July 2026, so a copy
  that still asks for them may already fail.
• The agent uses the Conversational Agent type of n8n's older Agent node, version 1.7. It has no tools.
• deepseek-r1 is a reasoning model, so a reply can begin with its chain of thought.
• With thinking on, the reasoning comes back in a field named reasoning_content beside content. Read the answer from
  choices[0].message.content.
• Billing: DeepSeek bills each API call, and the local model costs your own hardware.
• Your messages go to DeepSeek in the API examples and stay with your Ollama server in the local one.
• The chat trigger's public page is off until you turn it on in the trigger. Leave it off unless you want strangers
  using your keys.