Voices
Palmer built Kev with Devin and a pointer head rather than distilling Jev's outputs
The Kev repo credits Devin as a co-builder and states the models score options against decision tokens with a pointer head instead of training on Jev outputs, so it is an independent reimplementation of the System One idea rather than a distillation. Palmer says on X that the Qwen3.5 switch left the model "a bit overconfident" and regressed local performance, with another calibration run and MLX support queued between meetings. The whole port cost roughly $95 in rented GPU time, which is the actual builder takeaway: a scoped model plus a coding agent now compresses model development to an afternoon.
↳ Follow the thread