TatuEngine
03 de ago. de 2026
Operation Convergence — when TatuEngine decided to teach itself
After learning to infer (252× GPU speedup), TatuEngine went for the Master-Apprentice cycle: a Qwen2.5-3B teacher with LoRA distilling structured reasoning ([THOUGHT]/[ANSWER]) into the BitMamba-2 1B student. Six versions in one day (v0.17→v0.22.2), 381 examples, and the convergence that wouldn't come until the system message fix.
Continuar lendo →