The Chinese AI startup DeepSeek is rumored to be working on a new AI model called DeepSeek R2, following the success of its predecessor, R1. The new model is expected to combine high performance with extremely low costs, potentially shaking up the AI landscape.
The R2’s Architecture: A Hybrid Approach
Sources suggest that DeepSeek R2 will use a hybrid architecture based on Mixture of Experts (MoE). This could mean combining dense layers with advanced expert selection systems. The result? A model with around 1.2 trillion parameters, double the capacity of R1. This puts it on par with heavy hitters like GPT-4 Turbo and Gemini 2.0 Pro.
A Game-Changer in Cost Efficiency
The real surprise lies in the R2’s operational costs. Estimates suggest it’s up to 97% cheaper per token generated than GPT-4. This makes DeepSeek R2 a highly competitive option for businesses looking for scalable and affordable AI solutions. To put this into perspective, here are some key stats:
- Cost per token: up to 97% lower than GPT-4
- Training infrastructure: primarily Huawei’s Ascend 910B chips
- Efficiency: 82%
- Capacity: 512 PetaFLOPS in FP16 precision
A Shift Towards Local Technology
The use of Huawei’s Ascend 910B chips signals a strategic move by DeepSeek towards local technology to gain independence from Western supply chains. This could be a key factor in the company’s success, especially in a market where development costs are becoming increasingly important.
What’s Next?
While these rumors are yet to be officially confirmed, experts are already speculating about the potential impact of DeepSeek R2 on the AI market. If true, this could be a significant turning point, especially at a time when cost is a major deciding factor. According to WCCFTech, the tech community is eagerly awaiting official confirmation from DeepSeek.