- Google's New Gemini Lineup - Experimental
- Google has released experimental versions of Gemini 1.5 Pro, Flash, and Flash 8B, with Pro and Flash showing improved performance over previous versions. Notably, Flash 8B is a lightweight model that delivers satisfactory results on certain tasks.
Gemini 1.5 Flash 8b
Gemini Flash 8b was recently released on AI Studio.
First, for a while (until October 14th), there are no charges. (Even after that, a free tier exists, but with a maximum of 15 calls per minute and 1500 calls per day)
Currently, up to 4,000 calls per minute are provided free of charge. So, we are applying it to some services to test it, and we have conducted various performance tests.
First, performance.
Compared to Gemini Flash-002, the performance is definitely lower. It seems similar to the older Flash-001.
Previous PostIn the previous post, I mentioned that Flash8b is similar to Gemini Flash 001, and after actually using it, that's true.
The announced price is half that of Flash compared to Flash-8b, so I'm thinking about it a bit now. Can I keep using this...?
It seems that it can only be used for very simple functions. For example, it seems usable only for simple classification tasks. There are some drawbacks when complex tasks requiring prior knowledge need to be given to the LLM.
Speed.
According to the "announcement," the speed is faster compared to Flash, but I'm not sure. It's about the same, to the point where it's indistinguishable.
I haven't tried calling it 4,000 times per second, so I don't know the speed. (I don't think I'll ever use it that way.)
Using AI Studio, there are some safety filter issues.
When classifying news content, the safety filter was not released, causing occasional errors.
Summary.
For now, using only AI Studio makes it a little difficult to achieve overall desired usage. I'll have to test it again once it's introduced to Vertex AI.
Comments0