Ollama Streaming Responses: NDJSON Chunks, stream: false, and Handling Thinking and Tool Calls

By Web Asset House Studio · October 29, 2026

Note: Some future links on this site may be affiliate links. We never claim hands-on testing; guides are researched from the public sources listed in each article.

How we wrote this: A summary of Ollama's official Streaming (API reference), Streaming (capabilities) and Errors pages (linked below), checked October 29, 2026.

When you chat with a model in Ollama's app, words appear as they're generated. That's streaming. If you build on Ollama's local API, you choose whether to receive the answer piece by piece or all at once, and the defaults differ depending on how you call it.

The defaults

The docs give one rule that surprises people:

• REST API: streaming is on by default for endpoints that support it, such as /api/generate.
• Python and JavaScript libraries: streaming is off by default. Set stream=True (Python) or stream: true (JavaScript) to turn it on.

So a curl request that "returns a pile of JSON lines" and a library call that "returns one object" are both behaving as designed.

What a stream looks like

Streamed responses use newline-delimited JSON (content type application/x-ndjson): one JSON object per line, each carrying a small piece of the answer. In the docs' /api/generate example, each line has a response field with a token or two and "done": false. The last line has "done": true and a done_reason, such as "stop".

To rebuild the full answer, read line by line, parse each one, and append the text until you see done: true. With /api/chat, the text is in message.content instead of response.

Turning it off

Add "stream": false to the request body for any endpoint that supports streaming. The docs say you then get a single application/json object containing the complete text, with done: true.

Which to use

The docs' own guidance:

A chat interface almost always wants streaming. A script that classifies text or extracts JSON is usually simpler without it.

Streaming with thinking and tool calls

The capabilities page covers what else can show up in chunks:

• Chatting: each chunk has partial assistant content you can display as it arrives.
• Thinking: thinking-capable models send a thinking field alongside regular content. Watch for it to show or hide the reasoning before the answer.
• Tool calling: watch for tool_calls in chunks, run the requested tool, and add the result back into the conversation. See tool calling basics.

The docs stress one thing: accumulate the partial fields. Add up the thinking text and the content text as they arrive, then append a single assistant message with both to your message history for the next request. For tool calling this matters even more, since the model's thinking, its tool call and the tool result all need to go back in the next request.

The docs' Python example loops over the stream, prints chunk.message.thinking under a "Thinking:" label and chunk.message.content under "Answer:", and adds each to a running string. The Python library guide and JavaScript library guide cover setup.

Errors in the middle of a stream

Ollama's Errors page explains a subtlety. Normally, an error comes back with an HTTP status such as 400 (bad request), 404 (model not found), 429 (rate limit), 500 (server error) or 502 (cloud model unreachable), with a JSON body like {"error": "..."}.

But if something fails after streaming has started, the status code has already been sent and won't change. The error arrives as one more NDJSON line with an error property. So your stream reader should check every line for an error key, not just the HTTP status. Our troubleshooting guide shows where to look in the server logs when that happens.

Quick checklist

• REST: streaming on unless you send "stream": false.
• Libraries: streaming off unless you set stream to true.
• Read NDJSON line by line until done: true.
• Accumulate content and thinking; handle tool_calls.
• Check every line for an error field.

Sources

• Ollama docs: Streaming (API reference)
• Ollama docs: Streaming (capabilities)
• Ollama docs: Errors