www.analyticsdrift.com
Image source: Analytics Drift
On January 20, 2025, Chinese AI company DeepSeek launched the DeepSeek R1 large language model. This model is designed to mimic human reasoning and provide advanced analytical capabilities.
Image source: Deepseek
Compared to its competitors, OpenAI o1-mini, GPT-4o, and Claude 3.5, DeepSeek R1 has demonstrated superior performance in chemistry, coding, and mathematics.
Image source: Canva
The model has secured a 96.3 percentile rank compared to human participants, achieving exceptional benchmarks in MATH-500 (Pass@1) and GPQA Diamond (Pass@1).
Image source: Deepseek
This state-of-the-art model has a Mixture-of-Experts (MoE) architecture, similar to its company’s predecessor model, DeepSeek V-3—an open-source language model with 671 billion parameters.
Image source: Deepseek
The DeepSeek R1 model includes two versions, DeepSeek-R1-Zero and DeepSeek-R1. The former has only been trained in reinforcement learning without supervised fine-tuning.
Image source: Deepseek
Although the DeepSeek R1 model is not fully open-source, it offers significantly lower token prices than OpenAI. This makes the model a more affordable option for researchers and developers.
Image source: Deepseek