Home » News » DeepSeek R1: Open-Source AI Reasoning Model That Beats OpenAI’s o1

DeepSeek R1: Open-Source AI Reasoning Model That Beats OpenAI’s o1

by Ravi Teja KNTS
0 comment

DeepSeek released its V3 model last month. The company has now unveiled its reasoning model, DeepSeek R1. DeepSeek claims it not only matches OpenAI’s o1 model but also outperforms it, particularly in math-related questions. The good thing is that an R1 model is open-source, free to use, and can even run locally. Let’s explore if R1 is really that good.

What is DeepSeek R1?

DeepSeek R1 is a reasoning model, meaning it doesn’t simply provide the first answer it finds. Instead, it “thinks” through problems step by step, taking seconds or even minutes to reach a solution. This deliberate chain-of-thought process makes it far more accurate than traditional AI models and particularly useful in areas like math, physics, and coding, where reasoning is crucial.

DeepSeek achieves this reasoning capability through a combination of Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT). What? Here’s what these two terms mean:

  • Reinforcement Learning (RL): In RL, an agent learns by interacting with an environment and receiving rewards or penalties for its actions. The goal is to maximize the cumulative reward over time.
  • Supervised Fine-Tuning (SFT): SFT involves taking a pre-trained language model and further training it on a large dataset of high-quality text and code. This process refines the model’s abilities, improving its accuracy and performance on specific tasks.

Initially, DeepSeek relied solely on Reinforcement Learning without fine-tuning. This “DeepSeek R1 Zero” phase demonstrated impressive reasoning abilities, including self-verification, reflection, and generating long chains of thought. However, it faced challenges such as poor readability, repetition, and language mixing. To address these issues, DeepSeek combined RL with Supervised Fine-Tuning. This dual approach enables the model to refine its reasoning, learn from past mistakes, and deliver consistently better results. More importantly, this is an open-source model under the MIT License.

The Numbers Behind DeepSeek R1

DeepSeek R1 boasts a massive 671 billion parameters. Think of parameters as the brain cells an AI uses to learn from its training data. The more parameters a model has, the more detailed and nuanced its understanding. To put this into perspective, while OpenAI hasn’t disclosed the parameters for o1, experts estimate it at around 200 billion, making R1 significantly larger and potentially more powerful.

Despite its size, R1 only activates 37 billion parameters per token during processing. DeepSeek says it is done to ensure the model remains efficient without compromising reasoning capabilities.

The R1 model is built with the DeepSeek V3 model as its base, so the architecture and other stats are mostly similar. Here are the DeepSeek R1 model stats:

FeatureDeepSeek R1
ArchitectureTransformer with Mixture of Experts (MoE)
Total Parameters671 Billion
Activated Params37 Billion
Training Tokens14.8 Trillion
Context Window128,000 tokens
Output Limit8,000 tokens
Speed60 tokens per second
Open SourceYes

How Does R1 Compare to OpenAI’s o1?

When it comes to benchmarks, DeepSeek R1 is on par with OpenAI’s o1 model and even slightly surpasses it in areas like math. On math benchmarks like AIME, it scored 79.8%, slightly better than o1’s 79.2%. For programming tasks on Codeforces, it outperformed 96.3% of human programmers, showing it’s a serious contender. However, it’s slightly behind o1 in coding benchmarks.

For developers, the model is cheaper to integrate into apps. While the o1 model costs $15 per million input tokens and $60 per million output tokens, R1 costs just $0.14 per million input tokens (Cache Hit), $0.55 for million input tokens (Cache Miss) and $2.19 for output tokens, making it 90%-95% cheaper.

Another standout feature of R1 is that it shows its entire thought process during reasoning, unlike o1, which is often vague about how it arrives at solutions.

Distilled Versions for Local Use

DeepSeek has also released distilled models ranging from 1.5 billion to 70 billion parameters. These smaller models retain much of R1’s reasoning power but are lightweight enough to run even on a laptop.

Distilled Models:

Model NameBase ModelParameters
DeepSeek-R1-Distill-Qwen-1.5BQwen2.5-Math-1.5B1.5B
DeepSeek-R1-Distill-Qwen-7BQwen2.5-Math-7B7B
DeepSeek-R1-Distill-Llama-8BLlama-3.1-8B8B
DeepSeek-R1-Distill-Qwen-14BQwen2.5-14B14B
DeepSeek-R1-Distill-Qwen-32BQwen2.5-32B32B
DeepSeek-R1-Distill-Llama-70BLlama-3.3-70B-Instruct70B

These smaller models make it easy to test advanced AI capabilities locally without needing expensive servers. For example, 1.5B and 7B models can run on laptops. Whereas, 32B and 70B models deliver near R1-level performance but require more powerful setups. Even better, some of these models outperform OpenAI’s o1-mini on benchmarks.

Also Read:

How to Access DeepSeek R1

DeepSeek R1 is easy to access. Visit chat.deepseek.com and enable DeepThink mode to interact with the full 671-billion-parameter model.

Alternatively, you can access the Zero model or any distilled versions via the Hugging Face app, where you can download lightweight models to run locally on your computer.

Why DeepSeek R1 Matters

Outside of Microsoft’s Phi 4 model, there isn’t another open-source reasoning model available. Phi 4, however, has only 14 billion parameters and cannot compete with OpenAI’s o1 closed models. DeepSeek R1 provides a free, open-source alternative that rivals closed-source options like o1 and Gemini 2.0 Flash Thinking. For developers, the cost-effectiveness and open accessibility of R1 makes it especially appealing.

The only downside is that, as a Chinese-developed model, DeepSeek must comply with Chinese government regulations. This means it won’t respond to sensitive topics like Tiananmen Square or Taiwan’s independence, as the Cyberspace Administration of China (CAC) ensures that all responses align with “core socialist values.”

You may also like