Learn · Popular Open Source Models · The DeepSeek Breakthroughs
DeepSeek-R1: Open Reasoning Architecture
In plain English
DeepSeek-R1 is the open-source reasoning model that thinks before answering. It shows its inner thought process with <think> tags, catching its own mistakes just like OpenAI's o1! 🧠🔍
What is actually going on
DeepSeek-R1 utilizes large-scale reinforcement learning (RL) without supervised fine-tuning warmstart (R1-Zero) and multi-stage distillation. During inference, it allocates dynamic test-time compute to generate chain-of-thought reflection tokens. Distilled versions (1.5B, 7B, 8B, 14B, 32B, 70B based on Qwen and Llama) allow running deep reasoning on modest local GPUs.
Why it matters
DeepSeek proved that reasoning capabilities can be distilled into compact 7B-32B local models, making complex reasoning available on consumer laptops.
Another way to think about it
A detective who writes down all the clues and tests hypotheses in a notebook before announcing the culprit. 🕵️♂️
Check yourself
What special token tag does DeepSeek-R1 use to encapsulate its internal reasoning?
- <think>
- <code_block>
- <reason>
- <secret>
Show answer
<think> — DeepSeek-R1 wraps its step-by-step reasoning tokens inside <think> ... </think> blocks.
DeepSeek-R1 reasoning capabilities are available in compact distilled models (e.g. 7B and 14B).
- True
- False
Show answer
True — DeepSeek open-sourced distilled versions based on Qwen 2.5 and Llama 3 ranging from 1.5B to 70B parameters.