01 / Model
Wren 100: our second from-scratch model, in English and Korean
A 97.9M-parameter model pre-trained on 0.8 billion tokens of English and Korean and tuned on 340 million tokens of conversation data. It is open to chat now.
Wren 100 is the newest Wren: a dense transformer of about 97.9 million parameters, more than twice the size of Wren 40, trained from scratch rather than fine-tuned from someone else's weights.
It is pre-trained on 0.8 billion tokens of English and Korean text, then tuned on about 340 million tokens of conversation data. Compared with Wren 40 it keeps track of the conversation better and handles Korean.
It runs in your browser at ai.avenpa.com, behind Wren's effort setting: levels 3 to 5 use Wren 100 and levels 1 and 2 use Wren 40. It is still a small model. It can be wrong about facts, so check anything important.