Tuan's Blog

I’ve been using Hermes Agent for a couple of months now, testing out many different open-weight models like MiniMax M2.7 and MiMo-V2.5. I really love MiniMax M2.7, it runs everything without a single hiccup or problem. The only small drawback is the price, though it's already quite good compared to other SOTA (state-of-the-art) models.

But everything changed the day DeepSeek V4 came out. I’ve been using it extensively, especially DeepSeek V4 Flash. At first, I used the free version via my Nous Research subscription and OpenRouter, but now I’m trying it out directly on DeepSeek’s own platform.

So far, I have run 151 million tokens at a total cost of just $1.17. This absolutely blows my mind! Across those millions of tokens, the model follows instructions closely and with incredible accuracy. With Hermes Agent acting as the brain, I can just ask it to do something and know the outcome will be 90%+ correct.

The competition is fierce right now, and I am absolutely here for it. I really hope American AI companies start focusing heavily on efficiency to drive costs down like this.

-Tuan