The existing versions have less than half the parameters, but the performance is higher, so expectations were very high.
Since it's an Olllama cloud API and the server is working hard to generate tokens, it's fast.
It's incredibly fast in chat mode.
I tried using it by connecting OpenClok and Hermes open code Pi.
I'm very satisfied with the open code and Pi to this extent. You just have to give it a task.
The speed is good and the results are satisfactory,
but the problem is the agent tool, OpenClok and Hermes.
I usually have conversations with these agent tools and give them tasks,
but I have too many thoughts, so even if I turn off my thinking and proceed, it's the same.
Reducing the parameters and increasing performance seems to be a result of thinking deeply over and over again.
So, when I give it a task, it thinks a lot by itself and is slow. The results are good, but sometimes I don't know if it's working or stopped.
Also, even though I wrote in agent.md and memory.md to only use Korean, English, numbers, and special characters, it uses a lot of English after doing one task. When I asked why it keeps using English and to answer in Korean, it checked the log several times and made an excuse that its thoughts were just included in the message.
This seems to be a problem with the model itself.
These days, glm5.3-flash, qwen3.8-27b, and qwen3.8-flash-next are constantly being released, which is great.
