DeepSeek-R1 Repeats Itself or Outputs Gibberish: The 4 Settings That Fixed It
Endless loops, repeated sentences or mojibake from DeepSeek-R1 are almost always sampling or quantization — not a broken model. The exact parameters I now use.
DeepSeek-R1 Repeats Itself or Outputs Gibberish: The 4 Settings That Fixed It
Two failure modes, one root cause family. Either R1 repeats the same sentence until it hits the token limit, or it produces plausible-looking nonsense and broken characters. In every case I have debugged, it was sampling parameters, a bad quantization, or a chat template mismatch — never a "broken download" (that fails loudly instead).
Tested on deepseek-r1:8b and :14b via Ollama, 32 GB / RTX 3060 and Mac mini M4 16 GB.
Fix 1 — Use DeepSeek''s own recommended sampling
Related: GPT-5 Is Here: Everything You Need to Know About OpenAI's Most Powerful Model Yet →
Related: Ollama Pull Fails with "max retries exceeded" or EOF: How I Get Downloads to Finish →
The Ollama defaults are tuned for general chat and are too greedy for a reasoning model. What DeepSeek recommends, and what stopped the loops for me:
/set parameter temperature 0.6
/set parameter top_p 0.95
/set parameter repeat_penalty 1.1Temperature 0 is the classic trap: near-deterministic decoding is exactly what makes a model lock into a repeating cycle. I had it at 0.1 "for accuracy" and caused my own bug.
Persist it in a Modelfile:
FROM deepseek-r1:8b
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.1
PARAMETER num_ctx 8192ollama create r1-tuned -f ModelfileFix 2 — Stop over-quantizing
Related: Will AI Coding Agents Replace Developers? We Asked 100 Engineers →
Related: Ollama Filled My Disk: How I Moved the Models Directory Safely →
q2_K and q3_K_S on an 8B model produced word-salad for me about one response in five. q4_K_M is the floor I trust:
ollama pull deepseek-r1:8b-qwen-distill-q4_K_MIf memory is why you went below q4, shrink context or step down a model size instead — see the memory error fix.
Fix 3 — Do not paste your own chat template
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 →
R1 distills ship with a template that already inserts the reasoning scaffolding. Hand-writing <|User|> markers into the prompt made my output degrade badly. Send plain messages and let the runner apply the template.
Fix 4 — Watch for context overflow
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 →
Once the conversation exceeds num_ctx, the oldest tokens fall off — including the instruction — and the model starts drifting and repeating. Long research chats degrade for this reason alone.
/set parameter num_ctx 8192
/clear # start a fresh conversation instead of fighting a poisoned oneMy exact working profile
Related: Google Gemini 3 Ultra Review: Has Google Finally Caught Up? →
Related: Ollama Running Slow? 7 Fixes That Actually Worked on My Machine →
| Parameter | Value | Why |
|---|---|---|
| temperature | 0.6 | Below 0.3 loops on R1 |
| top_p | 0.95 | DeepSeek''s own recommendation |
| repeat_penalty | 1.1 | Above 1.3 broke coherence |
| num_ctx | 8192 | Room for the reasoning block |
| quantization | q4_K_M | q3 and below produced gibberish |
FAQ
Related: Midjourney vs DALL-E 4 vs Flux 1.1: The Definitive AI Image Generator Comparison →
Is repeat_penalty 1.5 safer? No. High penalties force the model away from words it legitimately needs and produce a different kind of nonsense.
Does the reasoning block cause the loop? No, but it consumes context fast, which triggers Fix 4 sooner than you expect.
Related reading Setup: DeepSeek-R1 offline on Mac mini M4. Reasoning tags: stripping think tags. Wider tooling: 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 Repeats Itself or Outputs Gibberish: The 4 Settings That Fixed It
Error I hitTwo failure modes, one root cause family. Either R1 repeats the same sentence until it hits the token limit, or it produces plausible-looking nonsense and broken characters. In every case I…
Exactly how I fixed itTested on deepseek-r1:8b and :14b via [Ollama](/article/ollama-common-errors-solutions-ar), 32 GB / RTX 3060 and Mac mini M4 16 GB.
- 2Step I tried
Two failure modes, one root cause family. Either R1 repeats the same sentence until it hits the token limit, or it produces plausible-looking nonsense and broken characters. In every case I…
Error I hitTested on deepseek-r1:8b and :14b via [Ollama](/article/ollama-common-errors-solutions-ar), 32 GB / RTX 3060 and Mac mini M4 16 GB.
Exactly how I fixed itRelated: [Ollama Pull Fails with "max retries exceeded" or EOF: How I Get Downloads to Finish →](/article/fix-ollama-pull-error-max-retries-exceeded)
- 3Step I tried
Tested on deepseek-r1:8b and :14b via [Ollama](/article/ollama-common-errors-solutions-ar), 32 GB / RTX 3060 and Mac mini M4 16 GB.
Error I hitRelated: [Ollama Pull Fails with "max retries exceeded" or EOF: How I Get Downloads to Finish →](/article/fix-ollama-pull-error-max-retries-exceeded)
Exactly how I fixed itRelated: [Ollama Pull Fails with "max retries exceeded" or EOF: How I Get Downloads to Finish →](/article/fix-ollama-pull-error-max-retries-exceeded)
- 4Step I tried
q2K and q3KS on an 8B model produced word-salad for me about one response in five. q4KM is the floor I trust:
Error I hitIf memory is why you went below q4, shrink context or step down a model size instead — see the [memory error fix](/article/fix-ollama-model-requires-more-system-memory).
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)
After all of the above, the setup that actually held up is the one described in this guide to “DeepSeek-R1 Repeats Itself or Outputs Gibberish: The 4 Settings That Fixed It”. 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.