Open-sourcing a full fine-tuning pipeline for embedded engineering — training toolkit + 35-domain MoE-LoRA model

TL;DR AI
2 min readKey summary
L'Électron Rare open-sourced two Apache 2.0 projects: KIKI-Mac_tunner, an MLX-based training toolkit for Mac Studio, and micro-kiki-v3, a routed embedded-engineering model.
micro-kiki-v3 is built on Qwen3.5-35B-A3B and uses a domain router plus 35 domain-specific LoRA stacks to handle specialized engineering tasks.
The release also includes supporting pieces for conflict resolution, bias reduction, memory, and a published dataset/pipeline with real engineering sessions and filtered open-source sources.
The goal is to reduce hallucinations and improve accuracy in embedded systems and electronics workflows, where general-purpose LLMs often struggle.

