Glossary

Short definitions for terms used in this book.

• QAT (Quantization-Aware Training): the model trained with low precision in mind, so the small 4-bit form keeps more skill than plain post-training shrinking.
• MoE (Mixture-of-Experts): a model with many expert subnetworks where only some fire per token. gemma4:26b-a4b-it-qat holds 26B parameters but only 4B activate per token, which buys speed.
• Dense: a model where all parameters work on each token. Simpler behavior, higher cost per token than MoE at the same size.
• Active parameters: the parameters that compute a given token. The cost driver for speed, not the total count.
• Context length: how much text (code, logs, prompts) the model can hold at once. Set via OLLAMA_CONTEXT_LENGTH in this book. Small values corrupt file writes. Large values cost RAM.
• Skill: a Markdown file under .pi/agent/skills/ with tool invariants and rules that little-coder loads into the system prompt.
• Headless: a command that runs to completion with no keyboard input, no REPL, and a clean exit. The opposite of interactive.
• Write-and-Notify: the strategy from write-temp-strategy.md: the agent writes full code to new tmp_* files and tells you the path, and you merge by hand.
• -it-: instruction trained. A model tuned to follow orders, not a raw base model.