Track 40Pro Track~135 min total
Build Your Own Private AI Lab
Run frontier open-weight models on your own hardware: chat, coding, images, video, and voice with zero per-token fees and full data privacy. Covers the hardware buying decision (NVIDIA, Apple Silicon, and AMD, with NVIDIA as the default), llama.cpp on CUDA, Metal, or Vulkan, picking and quantizing GGUF models, a ChatGPT-style UI with Open WebUI, local image/video with ComfyUI, voice cloning, wiring local models into agents and RAG, and the honest call on when local beats cloud. Built and benchmarked on real 2026 hardware, including the AMD ROCm gotchas nobody else documents.
Prerequisite:
§ The modules
- 01Why Run AI Locally (and What Hardware You Actually Need)The real case for local AI: privacy, zero per-token cost, no rate limits, full control. The three hardware paths (NVIDIA as the default, Apple Silicon, AMD), the memory-and-bandwidth math that decides everything, and named build recommendations for June 2026.Account
- 02Your Inference Engine: llama.cppWhy llama.cpp is the foundation, choosing your backend (CUDA for NVIDIA, Metal for Apple, Vulkan for AMD), installing on Windows, running your first model, and exposing an OpenAI-compatible API with llama-server.Account
- 03Choosing and Quantizing ModelsThe open-weight model landscape, GGUF quantization explained (why Q4_K_M is the sweet spot), the memory budget worksheet, dense vs MoE, and where to download trustworthy quants.Pro
- 04A ChatGPT-Style Interface with Open WebUIInstall Open WebUI the way that actually works (the Python 3.12 gotcha, pinning your data drive), connect it to llama.cpp or Ollama, and unlock multi-model chat, document RAG, and web search.Pro
- 05Local Image and Video Generation with ComfyUIGenerate images and video for $0/clip. ComfyUI setup, moving giant model libraries, Z-Image vs FLUX.2, Wan video, realistic generation times, and the AMD fp8 gotchas that crash the GPU.Pro
- 06Local Voice: TTS and Voice CloningAdd narration and a cloned voice to your stack with Qwen3-TTS running locally. Setup, the voice cloning workflow, building a fully local content pipeline, and the honest quality gap versus cloud TTS.Pro
- 07Wire Local Models into Agents, RAG, and Your WorkflowsThe OpenAI-compatible API as a universal connector, hybrid local/cloud routing, a private RAG second brain, plugging local models into coding tools and n8n, and the real cost comparison.Pro
- 08Going Further: Linux, vLLM, Fine-Tuning, and When NOT to Go LocalWhy Linux unlocks vLLM, official ROCm, and fine-tuning, the dual-boot vs WSL decision, serving at throughput with vLLM, a LoRA fine-tuning primer, and the honest framework for when cloud still wins.Pro
Start module 01 →Finish all six for the Pro Track