Which route is for you? A map of the Local LLM Inference course
The course has nine tracks and about 50 lessons. This map shows four routes by goal (just run models, build apps, tune performance, serve to others), what to read in what order, and what you can skip.
The course has nine tracks. You do not need all of them, and the order on the course page is not the only order. Pick the route that matches your goal.
#The tracks
| Track | What it covers | Level |
|---|---|---|
| Foundations | Tokens, context, prefill versus decode, sampling | Beginner |
| Hardware and Sizing | VRAM, memory bandwidth, Apple Silicon, buying | Beginner |
| Models and Quantization | GGUF, K-quants, AWQ, GPTQ, KV cache | Intermediate |
| Runtimes | Ollama, llama.cpp, vLLM, MLX | Intermediate |
| Performance Tuning | Offload, batching, caching, speculation, tuning playbook | Advanced |
| Serving and APIs | OpenAI-compatible API, Docker, load testing | Advanced |
| Building Applications | RAG, structured output, tools, agents | Intermediate |
| Edge and Browser | Small models, WebGPU, mobile | Intermediate |
| Evaluation and Operations | Speed, quality, licences, choosing models | Intermediate |
#Four routes
#1. "I just want to run models locally"
- 1Quick startrun a model in ten minutes
- 2Foundationsthe mental model
- 3Hardware and Sizingwhat fits on your machine
- 4Models and Quantizationpick a quant
- 5Runtimeschoose and use a runtime
Then dip into Evaluation and Operations when you need to choose a model.
#2. "I want to build applications"
- 1Quick start
- 2Foundations
- 3RuntimesOllama first
- 4Building ApplicationsRAG, structured output, tools, agents
- 5Evaluation and Operationsmeasure quality
Add Models and Quantization when you hit memory limits.
#3. "I want it fast"
- 1Foundations
- 2Hardware and Sizingbandwidth is the limit
- 3Models and Quantization
- 4Runtimes
- 5Performance Tuningincluding the tuning playbook
Continue to Inference Engineering for the maths behind it.
#4. "I want to serve it to other people"
- 1Foundations
- 2RuntimesvLLM especially
- 3Serving and APIsthe API, Docker, load tests
- 4Performance Tuningslots, batching, caching
Continue to Serving AI in Production for scaling and operations.
#Time
| Route | Approximate time |
|---|---|
| Run models locally | a few hours |
| Build applications | about a day or two |
| Make it fast | about a day or two |
| Serve to others | about a day or two |
Doing everything takes much longer; the routes exist so you do not have to. Free lessons are marked on the course page.
#What you can skip
- Hardware you already know? Skim VRAM math and run the calculators.
- Not on Apple Silicon? Skip the Apple lessons.
- Not building agents? Stop Building Applications after RAG and structured output.
#Exercise
Pick your route and write down the first three lessons you will do, and the one question you want answered by the end.
Answer: for example, route 1 with "will a 14B model fit on my 16 GB laptop?" The first three are the quick start, VRAM math and quantization basics.
Quick quiz
+10 XP per correct answer, +25 XP for a perfect score
1. You want to chat privately with a model on your laptop as soon as possible. Where do you go?
2. You want to serve a model to your team. Which tracks matter most?
3. Why read Foundations even if you are experienced?
Found this useful?
Send it to someone who is learning the same thing.