Chiisai: the models that will put Morfeul on the phone

Today, talking to Morfeul needs a connection. Our goal is for it to work anywhere, with no internet at all, and for that the model has to run on the phone itself. Chiisai is the family of compact models we are training for that job.

Small and specialised

A phone has a fraction of the memory and power of a data center, so Chiisai models are small, and compressed further to fit. That works because the job is focused: understand what someone on the farm is asking for, and turn it into the right action in Morfeul, such as starting a record, filling in its details or looking something up.

Trained on how people really talk

The training examples are requests written the way people on farms actually say them, each paired with what Morfeul should do. That is why we ask for examples on our training page: the more natural the requests, the better Morfeul understands people.

Tested on what it has never seen

Some examples are set aside and never used for training. Each model is judged on those, so we measure whether it understands new requests rather than whether it remembers old ones. Every compressed version is checked for answers Morfeul can read, the right action with the right details, replies short enough for a phone, and speed.

Then the phone

Only a version that passes those checks moves on to testing on real phones, where memory, speed and battery decide what is practical. That stage is next, and we will share the results as they come in, including where the models fall short.