Morgan Stanley has released a report highlighting the forthcoming release of DeepSeek's new AI language model, R2. This model boasts a significant upgrade, with 1.2 trillion parameters—more than double its predecessor R1. Notably, its active parameters have increased from 37 billion to 78 billion, enhancing language comprehension and output quality. The design aligns with top-tier models like Google's Gemini (GOOGL, Financial) and Anthropic Claude, aiming for more accurate AI cognitive abilities.
R2 promises to be both powerful and cost-efficient. The input cost is just $0.07 per million tokens, and output cost is $0.27, marking a dramatic cost reduction compared to OpenAI's GPT-4, which costs $2.50 and $10.00 for input and output, respectively. This cost advantage makes R2 an attractive option for AI developers, SaaS platforms, and corporate users.
Technologically, R2 offers improved multilingual and code generation capabilities, better multimodal abilities, and employs a Mixture of Experts (MoE) architecture for enhanced efficiency and reduced computing costs. Additionally, it reduces dependency on NVIDIA's (NVDA) H100 chips by using Huawei's Ascend 910B chips, marking a significant step towards China's AI tech independence.
