2026/Free online book · Points of View
In Defense of Small Models
Points of View, Volume VI
Access
Free
Chapters
11
Read time
113 min
Alpesh Nakrani
VP of Growth at ViitorCloud · former CTO & COO
The frontier gets the headlines; the small model gets the margin. Why most production value lives below the state of the art.
This edition is free to read onsite. Each chapter has its own URL, so readers can bookmark, share, and return to the exact section they need.
Table of contents
INTIntroduction: The Demo Was a LieWhy the largest model that wins the demo so often loses the production it was supposed to win, and what this book asks you to measure instead.9 min01What Small Actually MeansSmallness is not a parameter count; it is a property of cost, latency, memory, and control, and the families that have it look nothing alike.9 min02Capability Waste in Production AIYou are paying for intelligence your tasks never use; here is how to measure the waste before you decide what to do about it.9 min03Making the Task SmallerThe highest-use move is not picking a smaller model but shrinking the problem until a smaller model is obviously enough.9 min04The Workhorse Tasks: Classify, Extract, Route, SearchThe bulk of production AI is four boring jobs, and a small model or no model at all does each of them better than the frontier.9 min05Adapting Small Models: Fine-Tuning and DistillationHow to teach a small model your domain cheaply, and how to compress a strong teacher into a student that ships.9 min06Quantization and the Cost of CompressionHow to shrink a model's footprint and bill without quietly shrinking its quality, and how to prove you did not.8 min07Latency as a Product FeatureSpeed is not a constraint you tolerate; it is something users feel, and small models let you design it on purpose.9 min08Privacy, Local Inference, and Data ResidencySometimes the constraint that decides the whole architecture is not cost or speed but where the data is allowed to go.8 min09Routing: Small First, Large When NeededThe mature architecture is not one model but a cascade that handles the easy majority cheaply and escalates only what earns it.9 min10Proving Good Enough With EvalsYou cannot defend choosing a small model without measuring it, and measuring it well is a discipline most teams skip.9 min11When Small Models Are WrongThe honest chapter: where small models fail, where the compression reflex becomes its own mistake, and how to know the difference.8 minENDConclusion: A Sign of Maturity, Not a Lack of AmbitionBuilding a right-sized model portfolio, the production checklist, and why choosing the smallest sufficient system is the senior move.8 min