DeepSeek V3: 671B MoE beats GPT-4 at 10x lower cost

Posted by Bharadwaja Rishi · 2024-12-07

DeepSeek V3 is a 671B mixture-of-experts that matches GPT-4 on most benchmarks. Training cost reportedly 10x lower. Open weights. The AI cost curve is collapsing faster than predicted.

Explore GitVibes