NOTES ON THE NEW DEEPSEEK V3

Summary from the TLDR newsletter

Deepseek's flagship model, v3, is a 607B Mixture-of-Experts model with 37B active parameters. The company achieved remarkable performance while keeping training costs surprisingly low. The model outperforms other alternatives at its price point. It excels at reasoning and math, surpassing GPT-4 and Claude 3.5 Sonnet.