Hey everyone, as usual wanted to update you on some of the R&D initiatives we are backing including those by ICP Hub Egypt. This is the Motoko Coder open source series, ranging from 7B to 235B all trained on our data using our own compute clusters, those models had 0 compile rate in Motoko, now they are at 96%(for reference that beats Claude opus). The smaller models can be run from a mac, we are going to start releasing them along with API access, it will take us sometime because some of the models the team had abliterated so it’s not responsible that we release them atm. The rationale for this is that we often rely on our own models to cut away from dependencies and have data protection. Some of those models btw was tried with deterministic replays up to 120k tokens and we were able to produce determinism from same machines but that’s a topic for another time.
6 Likes