DeepSeek R1: The Open-Source Reasoning Model That Beat o1

Updated
7 min read
DeepSeek R1: The Open-Source Reasoning Model That Beat o1
TL;DR

DeepSeek R1 achieves 79.8% on AIME 2024 (vs o1-mini at 63.6%) using pure reinforcement learning without supervised fine-tuning, and it's MIT licensed.

What Makes R1 Different

Most frontier models are trained with supervised fine-tuning (SFT) on human-labeled reasoning chains, then refined with RLHF. DeepSeek R1 took a different path: it trained reasoning capability using pure reinforcement learning (GRPO) without any SFT cold-start.

The model learns to reason through trial-and-error against verifiable rewards (math answers, code correctness) rather than imitating human-written chains of thought. The result is a reasoning style that sometimes looks alien but achieves remarkable accuracy on hard problems.

The full research paper is publicly available and worth reading for anyone interested in RL-based training.

Benchmark Results

Click any column to sort
4 rows
AIME 202479.8%63.6%74.4%
MATH-50097.3%90.0%96.4%
Codeforces96.3th %ile93.4th %ile96.6th %ile
MMLU90.8%85.2%91.8%

R1 ties or beats o1 on most benchmarks while being MIT licensed and available to run locally.

Architecture: MoE With 37B Active Parameters

R1 uses a 671B parameter Mixture of Experts architecture, but only 37B parameters are active per forward pass. This gives it frontier-level capacity while keeping inference costs closer to a 37B dense model.

Distilled variants (trained to imitate R1's reasoning traces using SFT) are available at 7B, 14B, 32B, and 70B - making reasoning capability accessible on consumer hardware.

API Access

The DeepSeek API is OpenAI-compatible and priced at $0.14 per million input tokens (cache hits: $0.014) and $2.19 per million output tokens - a fraction of o1's cost.

python
from openai import OpenAI

client = OpenAI(
    api_key="your-deepseek-api-key",
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=[
        {"role": "user", "content": "Prove that sqrt(2) is irrational."}
    ]
)

# R1 returns both thinking and answer
print(response.choices[0].message.reasoning_content)  # chain of thought
print(response.choices[0].message.content)            # final answer

Running Locally With Ollama

bash
# 7B distilled  -  runs on consumer GPU (8GB VRAM)
ollama pull deepseek-r1:7b
ollama run deepseek-r1:7b "What is the integral of x^2 from 0 to 3?"

# 32B distilled  -  runs on 2× 3090s
ollama pull deepseek-r1:32b

When to Use R1

R1 is best for tasks with verifiable correct answers: math, formal proofs, competitive programming, structured data extraction with strict schema constraints. For open-ended creative or conversational tasks, GPT-4o or Claude 3.5 Sonnet often produce more natural output.

Summary

DeepSeek R1 is a landmark in open-source AI: frontier reasoning capability, MIT license, and a published training methodology that's already influencing how the industry thinks about RL-based training. Download weights at HuggingFace or call the API at deepseek.com.

#deepseek#reasoning#r1#chain-of-thought#open-source

One AI engineering post, weekly

LLM benchmarks, prompt techniques, and token-cost breakdowns, delivered to your inbox.

What should we test or explain next?

Comments

Have a question or something to add? Sign in to join the discussion.

Mahmudul Haque Qudrati — CEO & ML Engineer at Pristren

Mahmudul Haque Qudrati

CEO & ML Engineer

Software architect, full-stack engineer, and machine learning practitioner. Founder & CEO of Pristren, CTO at Venture Kick-backed PIN Sport Application GmbH (Basel, Switzerland), Software Development Lead at DailyFrame GmbH (building Dimetra patient transport software), and creator of Zlyqor. Focused on custom workflow automation, scalable web and mobile applications, and production AI/ML systems.