DeepSeek R2: Are There Any Predictions on the Performance?

Since DeepSeek R2 does not yet have an official release date, its rumored performance has sparked considerable discussion. Thus, please do not believe everything. Just treating this content as an aspect of thinking.

Let's start.

1. Hybrid MoE 3.0 Architecture

This architecture is an advanced Mixture of Experts (MoE) architecture that is to improve efficiency and performance in large-scale AI by smartly controlling which experts are active and how resources are used.

It's reported that the parameter scale reaches 1.2 trillion but with 78B activated.

If DeepSeek R2 is really used, reasoning-related energy usage will be reduced by 60%.

2. Chip Adaptation

Some reports that DeepSeek R2 uses Huawei Ascend 910B for training, not NVIDIA A100.

Huawei Ascend 910B is a second-generation AI processor chip developed by Huawei, optimized for both training and inference workloads.

More, it's said that the Huawei chip chieves 91% of the efficiency of a same-scale A100.

3. Lightweight

8-bit quantization compression technology reduces model volume by 83%.

And this will make the deployment on smart home devices possible.