Media & Culture

DeepSeek R2 achieves 98% MMLU, rivals GPT-4 with 70% lower cost

Open-source model beats GPT-4 on MMLU, costs a fraction

Deep Dive

A Reddit user submitted a post with a link and comments.

Key Points
  • Achieves 98.0% on MMLU benchmark, outperforming GPT-4's 86.4%
  • 671B total parameters with only 37B active per token, reducing cost by ~70%
  • Supports 128K context window and runs in 4-bit quantization on single A100 GPU

Why It Matters

Open-source AI now matches (or beats) proprietary models at a fraction of the cost, democratizing advanced AI.