LLM Compression Comparison Compare 4 Formats: Original, MOTH, TOON, kablUI

MOTH (Machine Optimized Text Hierarchy)

Type: LLM-based compression
Best for: Technical specifications, API definitions, database schemas, system architecture
Compression: 70-90% reduction vs. verbose formats
Philosophy: Blueprint not specification - answers "what" and "why", not exhaustive "how"

Visit MOTH Project Try Mothify GPT

TOON (Token-Oriented Object Notation)

Type: Programmatic compression
Best for: Structured data, configuration files, any JSON/YAML that needs deterministic compression
Compression: 40-60% reduction vs. JSON
Philosophy: Rule-based, deterministic compression using YAML-style indentation + CSV-style tabular layout

Visit TOON Project

kablUI (UI-Centric Compression)

Type: LLM-based compression
Best for: User interfaces, component hierarchies, design systems, UI specifications
Compression: 60-85% reduction vs. verbose UI definitions
Philosophy: Optimized for UI/interface compression with @BLOCKS, §HEADERS, and [Elements] syntax

Visit kablUI Project

Token Visualization

Click the "Tokens" button in the header to visualize how text is tokenized. Each token is wrapped in a colored border, making it easy to see:

• Token boundaries: Where one token ends and another begins
• Token density: How many tokens are needed for each format
• Compression impact: Visually compare token counts across formats

Algorithm: Real GPT tokenization using gpt-tokenizer (BPE - Byte Pair Encoding)

Encoding: o200k_base (used by GPT-4o, GPT-4.1, o1, o3, and other modern OpenAI models)

This uses the same tokenization as OpenAI's official tiktoken library. Hover over tokens to see their token IDs.

Token Counting Method

Token counts displayed below each panel are calculated using real GPT tokenization:

• Library: gpt-tokenizer (JavaScript port of tiktoken)
• Encoding: o200k_base (GPT-4o, GPT-4.1, o1, o3, etc.)
• Accuracy: Same tokenization as OpenAI's official models
• Method: BPE (Byte Pair Encoding) algorithm

The token counts you see are exactly what these models would use, not approximations.

How to Use This Tool

1. Select an example from the dropdown menu
2. Compare the 4 formats side-by-side: Original → MOTH → TOON → kablUI
3. Check token counts and compression percentages below each panel (using real GPT tokenization)
4. Click "Tokens" in the header to visualize token boundaries with colored borders
5. Click "Wrap" in the header to toggle word wrapping for all panels
6. Use the copy button on each panel to copy that format to your clipboard

Try It With Your LLM

Want to evaluate which compression format works best for your use case? Copy any compressed example and paste it into your favorite LLM (ChatGPT, Claude, Gemini, etc.) and try:

• Ask questions: "What does this specification define?" or "Explain this UI structure"
• Generate code: "Create a React component from this kablUI definition" or "Generate SQL from this MOTH schema"
• Build prototypes: "Create a working prototype based on this compressed spec"
• Extend it: "Add authentication to this API definition" or "Add a dark mode toggle to this UI"
• Compare outputs: Try the same prompt with MOTH vs kablUI vs TOON to see which format the LLM understands best

💡 The compressed formats use far fewer tokens than the original, leaving more room in your context window for the LLM's response!

All formats are open source and free to use. Visit the project pages for documentation and implementation details.