NPC Fast 1.7B: Building a Usable Small Model on a Single H100

We describe NPC Fast 1.7B, a dual-purpose small language model trained on a single H100 to act as both a general-purpose assistant for simple queries and a routing classifier that escalates harder queries to a domain specialist. The model is built on SmolLM2-1.7B-Instruct via two stages: full-weight continual pretraining at 4K and 16K context using a curriculum schedule with YaRN RoPE configured for longer extension, followed by a small router-head LoRA fine-tuned on 500 synthetic prompt/label pairs in 96 seconds of wall-clock training. The router achieves 98.3% accuracy on a held-out out-of-distribution query set; the model passes 77.8% on IFEval and 100% on a small in-distribution function-calling sanity suite. We document the full pipeline that produced these numbers, including five concrete bugs that surfaced during development and the workarounds we applied. We trained only stages 1-2 of a planned five-stage curriculum, leaving the model's effective context at 16K despite YaRN being configured for 131,072. Needle-in-haystack retrieval at 32K is 0%, which we document rather than hide. The total compute footprint of the project is roughly 100 H100-hours. The contribution is not a novel method. It is a documented end-to-end recipe for building a serviceable small model with research-taste defaults on hardware accessible to an individual researcher, with the bugs and the unmet ambitions reported alongside the wins.