Key Points

  • DeepSeek R2 scores 92.4% on HumanEval+ vs GPT-5's 93.1% — a 0.7% gap at radically lower cost
  • Training cost estimated at $8.6M vs GPT-5's rumored $500M+, using Mixture-of-Experts architecture
  • R2 supports 256K context window with 4x faster inference than GPT-5 on long documents
  • Open-weight release expected under MIT license, enabling self-hosting and fine-tuning
  • Chinese government labs contributed novel quantization techniques that reduce VRAM requirements by 40%

Why It Matters

If R2 delivers on these benchmarks, it breaks the assumption that frontier AI requires billion-dollar training budgets. Open-weight models at GPT-5 performance mean startups can self-host competitive models, enterprise data never leaves private infrastructure, and the cost of AI inference drops dramatically. The MoE architecture also suggests a path to efficient models that don't require constant hardware upgrades.

Sources