Tina Huang
YouTuber who teaches AI to builders, with quizzes, summary slides and free guides, in a course-like format. Their Hermes agent, which runs much of their life and work, uses local Qwen3 models on their own hardware.
In Every Size Local AI In 24 Minutes (Sept 2026) they run local models on six hardware classes:
- ESP32-S3 (8 MB): TinyStories only. The Arduino Uno R4 (32 KB) fits nothing but can call a bigger machine over Wi-Fi.
- Raspberry Pi 5: a Whisper, Qwen3 1.7B and Piper voice assistant, plus Moondream vision.
- iPhone 16: Gemma 4, Stable Diffusion 1.5 and YOLO11n through apps.
- 32 GB MacBook Pro: every model category.
- 64 GB Mac Studio and 128 GB AMD Ryzen AI Halo: bigger models, always on.
- A rented RTX 4090 and 8x H100.
Their teaching device is a kitchen analogy for the hardware, and their lesson is capacity-vs-bandwidth. They also give a sizing rule (ram-to-parameters-rule).
Source: report
Linked from
Capacity vs bandwidthMemory-bandwidth-bound generationRAM-to-parameters rule