Local Models That Don't Feel Dumb

What to Run on 24, 64, and 128 GB, 2027 Edition

von Ravi Vale

182 Seiten, Taschenbuch

Kurzbeschreibung des Verlags

Your local model isn't dumb the way you think it is. It's dumb because of a setting nobody showed you, and a setting is something you can change tonight.There's a folder on your disk of models you tried once. Each one got a verdict: slow, forgetful, oddly polite about not doing the job. Every one of those verdicts was honest, and every one was formed on a file whose name you never read, at a conversation length something else chose, under a runner version you never wrote dowWhat you get, in fourteen chapters and three appendices: - The dumb test. Twelve prompts and a scoring sheet, ten minutes, run twice: once with the window nearly empty, once with a long document already in it. The gap between those two columns is the number nobody in your folder ever had. Chapter 1 hands you the sheet and asks you to run the first prompt before you read another page.- Three fit tables, in bytes. One for each memory size. Every entry is a filename in full, the byte count as the repository serves it to two decimals, what a conversation costs per token from the model's own configuration file, the context that fits in what's left, and the license as the card states it. No leaderboard score sits beside any file, and nothing was filled in from memory: where a score could not be read on the day, the page says NOT FOUND.- The six causes and their fixes. The window the runner picked; the model spilling onto the processor; the runner version that moved; the wrapper around your message; the tools that were never sent; the file cut too far. Two of them take one command to check and no download.- Quantization in one chapter, licenses in one chapter, and a last chapter on how to re-run every number in the book yourself in December, from the pages they came from.This book is for the reader who: - owns a 5090 tower, a Mac, a Strix Halo mini or a Spark, and has been disappointed twice- runs a coding agent against a local endpoint and has watched tool calls silently vanish- does not want a leaderboard, an architecture diagram, or a Discord's opinion- would rather take a measurement than form a feelingEvery parameter count, license, context length and file size comes from the model's own card or repository as read on 5 September 2026, and is cited. Every speed number carries its source, its method and its date. No pick prints as tested unless a machine and a date are named; every other pick prints the words sized, not tested, and means them. The Kindle edition is refreshed each quarter; the preface names the day the tables were read.Read chapter one free with Look Inside. Written against Ollama 0.33.3 and LM Studio 0.4.23: when this book and the model card disagree, the card is right.

• Reihe: Independently published
• ISBN: 9798176089752
• Sprache: Englisch
• Erscheinungsdatum: 21.09.2026
• Größe: 229 x 152 mm
• Verlag: Independently published