Small, local-first language models. Nemotron fine-tunes, reward models, deterministic routing, and GGUF exports. Everything runs on consumer GPUs.