Interactive visuals and materials for the Little Coder video (local coding agents with small LLMs).
Every demo is a single self-contained HTML file — open it directly in a browser, no build step or server needed. All pages support light and dark mode. Start at index.html for a linked hub of everything.
| File | What it shows |
|---|---|
terminal-bench.html |
Terminal-Bench 2.0 scores: Qwen 3.6 + Little Coder vs Gemini CLI combos vs the frontier. Animated bars with replay. |
model-comparison.html |
The two models used in the experiment — Gemma 4 31B QAT (dense) vs Qwen 3.6 35B-A3B (MoE) — and why active params set the speed. |
model-family.html |
Model-family explorer: brain sizes, file size per quant, memory needs, and what hardware each variant fits on. |
quantization.html |
Quantization explained via the "photo over Messenger" analogy, with a live-posterized image and real bits/weight math. |
can-i-run-it.html |
Hardware calculator: pick a machine, model size, quant, and context — get memory fit, tokens/sec estimates, and a verdict. |
hosting-flow.html |
Step-through of the local-model pipeline: download → RAM → context window → localhost server → Little Coder. |
Both models were given the same one-shot prompt to build a vacation-tracking SaaS. The exact prompt is in PROMPT.md.
Hardware: MacBook Pro, M3 Max, 64 GB unified memory, models served via LM Studio.