gpt2 vs DeepSeek-R1

Try Try

gpt2 has 137M parameters (137M active per token) and takes 548 MB on disk; DeepSeek-R1 has 684B (36.6B active) and takes 689 GB. Compare them layer by layer.

At a glance

openai-community/gpt2deepseek-ai/DeepSeek-R1
Repo createdMar 2, 2022updated Feb 19, 2024Jan 20, 2025updated Mar 27, 2025
Model typeDenseFull attention (MHA)Mixture of expertsFull attention (MLA)
InputsTextText
Total parameters137M4995× less684B4995× more
Active per token137M267× less100% of the model36.6B267× more5.4% of the model
ExpertsNone (dense)8 of 256 activeplus 1 shared, always on
Max context (from config)1K tokens160× less160K tokens160× more
Layers125.1× less615.1× moreplus 1 extra prediction layer
On disk548 MB1256× less1 file689 GB1256× more163 files
PrecisionFP32 (100%)FP8 E4M3 (99%), BF16 (1.1%)
QuantizationNoneoriginal precision (FP32)FP8 E4M3blocks of 128×128 · 99% of parameters
Fewest GPUs1× RTX 4060a single 8 GB card · weights only6× NVIDIA H200fits in one 8-GPU server · weights only
Licensemitmit
GitHubopenai/gpt-2deepseek-ai/DeepSeek-R1