DeepSeek-R1 Shows Its <think> Tags in the Output — Here Is How I Strip Them
DeepSeek-R1 prints its whole reasoning chain before the answer. Three ways I remove it: prompt-level, API-level parsing, and a small wrapper I use in every script.
DeepSeek-R1 Shows Its <think> Tags in the Output — Here Is How I Strip Them
DeepSeek-R1 is a reasoning model, so every response starts with a long block wrapped in think tags before the real answer. In a chat window it is interesting. In an automation that writes to a database, it is garbage in your column.
Tested on: deepseek-r1:8b and deepseek-r1:14b (q4_K_M) via Ollama 0.6.x, Mac mini M4 16 GB and an RTX 3060 box.
Why prompting alone does not work
Related: GPT-5 Is Here: Everything You Need to Know About OpenAI's Most Powerful Model Yet →
Related: DeepSeek-R1 Repeats Itself or Outputs Gibberish: The 4 Settings That Fixed It →
I tried "do not show your reasoning" in the system prompt on both sizes. R1 ignores it — the reasoning block is part of how the model was trained, not an instruction it can opt out of. Suppress it in code, not in the prompt.
Fix 1 — Read the dedicated reasoning field
Related: Will AI Coding Agents Replace Developers? We Asked 100 Engineers →
Related: Ollama Filled My Disk: How I Moved the Models Directory Safely →
Modern Ollama separates it for you, which is the cleanest path:
import ollama
r = ollama.chat(model="deepseek-r1:8b",
messages=[{"role": "user", "content": "Summarise this changelog"}])
print(r["message"]["content"]) # answer only
# r["message"].get("thinking") # reasoning, if you want to log itIf your client library is older, upgrade first — that alone removed the problem in two of my scripts.
Fix 2 — Strip it with a regex
Related: The 27 Best AI Tools in 2026 (Tested for 90 Days) →
Related: Ollama Answers Get Cut Off Mid-Sentence: num_ctx vs num_predict Explained →
The one-liner I keep in my utils file:
import re
def strip_reasoning(text: str) -> str:
return re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL).strip()Guard against a truncated block — if generation hits the token limit mid-reasoning, the closing tag never arrives:
if "<think>" in out and "</think>" not in out:
out = "" # incomplete; retry rather than store noiseFix 3 — Turn thinking off where the runner supports it
Related: ChatGPT vs Claude 4: Which AI Should You Actually Pay For in 2026? →
Related: Ollama "connection refused on 127.0.0.1:11434": The 5 Causes I Have Actually Hit →
ollama run deepseek-r1:8b --think=falseQuality on multi-step tasks drops noticeably when I do this, so I only use it for classification and short extraction jobs where R1 was overkill anyway.
My exact settings
Related: Google Gemini 3 Ultra Review: Has Google Finally Caught Up? →
Related: Ollama Error "model requires more system memory": How I Fixed It in 10 Minutes →
| Use case | Model | think | num_ctx |
|---|---|---|---|
| Long-form drafting | deepseek-r1:14b | on, stripped in code | 8192 |
| Tag/classify | qwen2.5:7b | n/a | 4096 |
| Offline research | deepseek-r1:8b | on, logged | 8192 |
Note the middle row: for pure classification I switched off R1 entirely. It was 3× slower for no measurable gain.
FAQ
Related: Midjourney vs DALL-E 4 vs Flux 1.1: The Definitive AI Image Generator Comparison →
Does the reasoning block count against my context?
Yes. It is generated tokens, so it eats both time and num_ctx. That is why my R1 sessions use a bigger context than my Qwen ones.
Can I show the reasoning in a UI?
That is exactly what I do — render thinking inside a collapsed "show reasoning" toggle, and keep the answer clean.
Where do I set R1 up from scratch? My full walkthrough is running DeepSeek-R1 offline on a Mac mini M4, with the general path in the local LLM guide and tooling context in the ultimate AI tools guide.
Experience Log
This is not theory. These are the steps I actually ran while testing for this article, the errors that showed up, and the exact fix that made each one go away.
- 1Step I tried
DeepSeek-R1 is a reasoning model, so every response starts with a long block wrapped in think tags before the real answer. In a chat window it is interesting. In an automation that writes…
Error I hitTested on: deepseek-r1:8b and deepseek-r1:14b (q4KM) via [Ollama](/article/ollama-common-errors-solutions-ar) 0.6.x, Mac mini M4 16 GB and an RTX 3060 box.
Exactly how I fixed itRelated: [DeepSeek-R1 Repeats Itself or Outputs Gibberish: The 4 Settings That Fixed It →](/article/fix-deepseek-r1-repeating-loops-gibberish)
- 2Step I tried
python import ollama r = ollama.chat(model="deepseek-r1:8b", messages=[{"role": "user", "content": "Summarise this changelog"}]) print(r["message"]["content"]) # answer only #…
Error I hitIf your client library is older, upgrade first — that alone removed the problem in two of my scripts.
Exactly how I fixed itRelated: [Ollama Answers Get Cut Off Mid-Sentence: numctx vs numpredict Explained →](/article/fix-ollama-answers-cut-off-num-ctx-num-predict)
- 3Step I tried
Quality on multi-step tasks drops noticeably when I do this, so I only use it for classification and short extraction jobs where R1 was overkill anyway.
Error I hitRelated: [Ollama Error "model requires more system memory": How I Fixed It in 10 Minutes →](/article/fix-ollama-model-requires-more-system-memory)
Exactly how I fixed itRelated: [Ollama Error "model requires more system memory": How I Fixed It in 10 Minutes →](/article/fix-ollama-model-requires-more-system-memory)
- 4Step I tried
| Use case | Model | think | numctx | |---|---|---|---| | Long-form drafting | deepseek-r1:14b | on, stripped in code | 8192 | | Tag/classify | qwen2.5:7b | n/a | 4096 | | Offline research…
Error I hitNote the middle row: for pure classification I switched off R1 entirely. It was 3× slower for no measurable gain.
Exactly how I fixed itNote the middle row: for pure classification I switched off R1 entirely. It was 3× slower for no measurable gain.
After all of the above, the setup that actually held up is the one described in this guide to “DeepSeek-R1 Shows Its <think> Tags in the Output — Here Is How I Strip Them”. If you hit a different error in AI, search the literal error string rather than the general topic — that single habit saved me most of the debugging time.
Related Articles
مقالات ذات صلة — تابع القراءة داخل الموقع
مواضيع مقترحة · Suggested Topics
استكشف مواضيع ومحاور ذات صلة بهذا المقال — روابط داخلية لتعميق قراءتك.
The Daily Pulse
Get the 5 biggest tech stories in your inbox every morning. Free, no spam, unsubscribe anytime.
Join 50,000+ tech professionals reading every day.