Summary
This tutorial walks through building a complete 3D animated commercial website using Qwen 3 27B running locally on a Mac Studio via MLX quantization (4-bit). The workflow integrates the Pi agent framework—a lightweight terminal-based alternative to Claude Code—with Higsfield AI for generating on-brand images and videos directly from the command line via a CLI tool, demonstrating a fully local-first AI development stack with cloud-based media generation.
The demonstration builds a roofing company website from scratch, covering the complete pipeline: generating a photorealistic hero image using Higsfield with specific camera angle and lighting prompts, scaffolding responsive HTML/CSS/JavaScript with hover effects and testimonial sections, producing a 10-second scroll-triggered video using Sea Dance 2.5, and applying design refinements to reduce the characteristic aesthetic artifacts common in locally-generated AI output. Qwen’s multimodal vision capabilities are used throughout to inspect generated assets and iterate without leaving the terminal environment.
The tutorial also demonstrates Pi’s skill system for extending agent capabilities with new tools (analogous to MCP), and covers practical workflow decisions like asset organization in VS Code and when to regenerate versus edit images. The result is a polished website with animated sections, video background effects, and service listings—illustrating what is achievable when pairing open-weight local models like Qwen 3 with specialized generative media APIs.
📺 Source: Bart Slodyczka · Published August 19, 2026
🏷️ Format: Tutorial Demo
