For years, the AI community has chased a moonshot: creating open-source models that rival the reasoning power of giants like OpenAI. Today, that moonshot just landed. DeepSeek-R1, a new open-source language model released under the MIT license, not only matches OpenAI’s cutting-edge “o1” models in reasoning benchmarks — it does so at a fraction of the cost. Let’s unpack why this matters and how DeepSeek pulled it off.

The DeepSeek Breakthrough: AI That Thinks Step-by-Step

DeepSeek-R1 is part of a new class of “thinking models” that mimic human-like reasoning. Unlike traditional language models that generate answers in a single pass, DeepSeek-R1 breaks problems down, debates alternatives, and self-corrects — all visible in its “Chain of Thought” outputs. For example, when asked “How many Rs are in ‘strawberry’?”, the model writes:

“First, I’ll spell it out: S-T-R-A-W-B-E-R-R-Y. Now I’ll count: positions 3 (R), 8 (R), and 9 (R). Wait, is that right? Let me check again… Yes, three R’s.”

This isn’t just a parlor trick. On benchmarks like AIME 2024 (a math competition), DeepSeek-R1 outperforms OpenAI o1, and it’s neck-and-neck on coding tasks (Codeforces) and real-world problem-solving (SWE-Bench). Even more impressive? It does this while being 10x cheaper than OpenAI’s API (0.14vs.0.14vs.15 per million tokens for outputs).

How They Built a “Thinking Machine”

The team tackled a critical problem: How do you teach an AI to reason without massive human feedback? Traditional methods rely on supervised fine-tuning (SFT), where humans manually craft examples. DeepSeek’s answer? Reinforcement Learning (RL) on steroids.

✅ DeepSeek-R1-Zero: The AlphaGo of Language Models The first model, R1-Zero, learned purely through trial and error using a technique called Group Relative Policy Optimization (GRPO). Here’s the twist:

• No supervised data: Unlike OpenAI’s o1, R1-Zero skipped SFT entirely. It started with raw pretraining data and learned by generating answers, comparing them in groups, and rewarding correct reasoning.
• Self-evolution: Over time, the model taught itself to spend more “thinking time” on harder problems. In one experiment, it generated 3x longer reasoning chains for complex math questions, without being told to do so.

✅ DeepSeek-R1: Fixing the Quirks

R1-Zero had flaws: its outputs were messy (mixing languages like English and Chinese) and hard to read. The team fixed this with a “cold start” phase:

• Mini-SFT: They fine-tuned the model on a tiny dataset (just 1,000 examples) of high-quality reasoning chains.
• Two-stage RL: First, they trained for accuracy and format. Then, they added a second RL stage to align with human preferences (e.g., helpfulness, safety).

The result? A model that thinks clearly, stays on-task, and even outperforms GPT-4o on coding benchmarks like LiveCodeBench.

The Secret Sauce: Technical Innovations

✅Group Relative Policy Optimization (GRPO)

Instead of using a separate “critic” model (like OpenAI’s PPO), GRPO compares multiple responses in a group. Analogy: Imagine students working on a math problem. The teacher rewards the group based on relative performance, not absolute scores. This pushes the model to self-improve competitively.

✅Reasoning-Oriented Rewards