Descriptions:
Craig Hewitt walks through a complete local LLM setup using two distinct approaches: LM Studio for a graphical interface and Ollama for terminal-based operation. The tutorial targets Apple Silicon users—specifically an M4 Mac Mini with 16GB of unified RAM—and begins by using canIrun.ai to determine which models the hardware can handle before downloading anything.
The video explains the full stack from the ground up: how models stored as GGUF files (sourced from Hugging Face or Ollama’s library) relate to inference servers, and how those connect to downstream tools like coding assistants (Continue, OpenClaw) or IDE integrations. Hewitt demos Gemma 4 inside LM Studio, highlighting its multimodal capabilities for text, image, and reasoning tasks, then runs the exact same prompt—designing a podcast hosting platform like his company Castos—through both Gemma 4 locally and Claude Opus 4.6 via Claude Code.
The comparison finds the outputs broadly comparable for everyday tasks, leading Hewitt to argue that local models deliver roughly 90% of what most users need, entirely free and private. The video also covers how to connect local inference servers to open-source coding tools as drop-in replacements for subscription-based services like Claude Code or Codex, making a practical case for local AI as a path to both cost savings and model sovereignty.
📺 Source: Craig Hewitt · Published April 06, 2026
🏷️ Format: Tutorial Demo
