Every AI model you have ever used — ChatGPT, Claude, Gemini — is someone else's dragon. Powerful, impressive, and completely under someone else's control. This course teaches you to raise your own: a language model that runs on your hardware, learns from your data, and answers in your voice. No API bills, no rate limits, no data leaving your machine.
The title is a wink at a certain movie about a boy and his dragon, because the metaphor genuinely fits. A base language model is a wild creature: enormously capable, but not yet yours. Training it — feeding it the right data, correcting its behavior, testing its flight — is how it becomes a companion instead of a curiosity.
What "Training Your Own Model" Actually Means in 2026
Let's be precise, because the words get abused. You will not be training a model from scratch. Training GPT-class models from zero costs tens of millions of dollars and a datacenter. What you will do is fine-tuning: taking a powerful open-weights model (Qwen, Llama, Gemma, Mistral) that already speaks fluent English, and teaching it your specific task, tone, and knowledge.
Concept
Base model = a dragon raised by someone else (Meta, Alibaba, Google, Mistral) and released into the wild with open weights. Fine-tuning = adopting that dragon and training it for your job: answering your customers, writing in your style, knowing your domain.This is not a toy workflow. In 2026, companies fine-tune small open models because a well-trained 4B model that knows your product beats a generic giant model at that one job — while running locally for free, with full data privacy.
What You Will Build in This Course
Across 8 lessons, you will train Ember — your own fine-tuned model. Concretely, you will:
- Pick a realistic training goal and define what "success" looks like
- Understand exactly what your hardware can and cannot train (we cover two real setups: a Mac Studio M3 Ultra with 96 GB unified memory and a PC with an Intel i9 and an RTX 4090 with 24 GB VRAM)
- Build a clean instruction dataset in JSONL format
- Run a real LoRA fine-tune with MLX (Mac) or Unsloth (PC)
- Evaluate the results like an engineer, not a fan
- Level up with QLoRA, preference tuning, and quantization
- Export your model to GGUF and serve it with Ollama or LM Studio, or share it on Hugging Face
Honest Note
Fine-tuning will not give your model new reasoning superpowers, and it will not reliably teach it facts. It changes behavior: tone, format, task discipline, domain vocabulary. If you need up-to-date facts, that's a job for RAG — and lesson 2 teaches you to tell the difference before you waste a week of training runs.Why Local Training Is Suddenly Realistic
Three things converged to make this course possible on a desk, not a datacenter:
1. Small open models got shockingly good. A 2026-era 4B–8B model outperforms the 70B models of two years ago on most practical tasks. Small enough to train at home, good enough to matter.
2. Parameter-efficient fine-tuning (LoRA/QLoRA). Instead of updating all billions of weights, you train tiny "adapter" matrices — often less than 1% of the model. This slashes memory needs by an order of magnitude. Your dragon learns new tricks without regrowing its whole brain.
3. Consumer hardware caught up. Apple's unified memory lets a Mac Studio hold huge models in RAM that the GPU can use directly. NVIDIA's RTX 4090 brings serious training compute for a fraction of datacenter prices. Both setups in this course are real machines people own, not hypotheticals.
The Two Machines We'll Use Throughout
Every hands-on lesson gives you two paths — follow the one matching your hardware:
- Path A — Mac Studio, M3 Ultra, 96 GB unified memory. Training runs on Apple's MLX framework. Huge memory means huge models fit; raw training speed is more modest.
- Path B — PC, Intel i9 + RTX 4090 (24 GB VRAM). Training runs on CUDA with Unsloth and Hugging Face libraries. Less memory, but brutal compute speed — the 4090 trains small models faster than the Mac.
You don't need these exact machines. Any Apple Silicon Mac with 32 GB+ or any NVIDIA GPU with 12 GB+ can follow along with smaller models. Lesson 3 gives you the exact memory math to know what fits on your machine.
Prerequisites
- Comfort with a terminal — you can open one,
cdaround, and run commands. No programming degree needed. - Python installed (3.10+) — we'll use
pipto install everything else. - Disk space — 50 GB free. Models are chunky.
- Patience for iteration — your first trained model will be mediocre. That's normal, and lesson 6 is dedicated to making it good.
Try It
Before the next lesson, install a local model runner and feel what "a model on your machine" is like. Download LM Studio (lmstudio.ai) or Ollama (ollama.com), pull a small model likeqwen3:4b, and chat with it. Notice what it does well — and where it feels generic. That generic feeling is exactly what training removes.