Michelangelo Romero Chisco

Fund GPU/TPU compute for Terse Mini (1.5B active / 4.5B total ternary sparse transformer) pre-training on 200B tokens. All weights and training code published openly under Apache 2.0.

Extend ternative to serve Terse models: MoE expert dispatch, KDA (Kimi Delta Attention) kernels, MLA with KV absorption, recurrent depth, Apple Metal backend. This makes ternative the reference inference engine for the next generation of ternary models.