Toonify: A Compact Data Exchange Format Optimized for LLMs, Saving Up to 60% Tokens

The Toonify project proposes the TOON compact data format, specifically designed for large language models (LLMs), aiming to address the token consumption cost issue in LLM usage. This format can save up to 60% of token usage while maintaining human readability, supports cross-platform use, and is of great value for LLM application scenarios that frequently transfer structured data (such as Agent systems, RAG, tool calls, etc.).

In the context of LLMs, JSON's redundant syntax (such as quotes, brackets, line breaks) increases token consumption. For example, a simple JSON object contains many syntax symbols, each of which is a token for the Tokenizer. The core design principles of TOON are: 1. Remove key-value quotes when there is no ambiguity; 2. Simplify separators to represent hierarchy and lists; 3. Reduce whitespace characters; 4. Maintain human readability, distinguishing it from binary formats.

Cross-platform Support: Available for Windows (.exe), macOS (.dmg), and Linux (compressed package + command line), with minimum requirements of a dual-core processor, 4GB RAM, and 200MB storage.

Core Features: Supports JSON/YAML import, TOON conversion, reverse export to JSON, and batch processing of folders.

Usage Flow: Open the application → Import to load files → Select output format → Convert → Save results; no code required, easy for non-technical users to get started.

LLM Tokenizers (such as GPT's BPE algorithm) treat punctuation, spaces, and line breaks as independent tokens. TOON saves tokens through the following strategies: 1. Remove JSON's syntax noise (e.g., tokens consumed by quotes); 2. Compact nested representation to reduce indentation and line break tokens; 3. Intelligent key-value separation to minimize token overhead of separators.

API Call Optimization: Suitable for Multi-Agent message passing, RAG context input, Function Calling parameter serialization, reducing call costs.

Local Model Acceleration: Optimize context window (accommodate more information), improve inference speed (reduce token preprocessing), and lower memory usage.

Data Storage and Transmission: Compact log records, configuration files, and API responses, reducing storage and bandwidth consumption.

1. Ecosystem Compatibility: Requires additional conversion steps to integrate with existing tools; debugging tools may not support it; teams need to learn new specifications. 2. Readability Trade-off: Over-compactness may increase the burden of human reading. 3. Uncertain Saving Ratio: Depends on data characteristics (key length, nesting depth, data type); savings for pure numerical data are limited.

• vs MessagePack: TOON maintains text readability, while MessagePack is binary and unreadable. - vs YAML: TOON has stronger parsing robustness and avoids YAML's indentation sensitivity issues. - vs Custom DSL: TOON is a general-purpose format, not specific to a particular domain.

Toonify reflects the demand for data format optimization in the LLM era and has direct engineering value in the context of token costs and scarce context. Although TOON will not replace JSON, it is a valuable option for high-frequency LLM interaction scenarios. As multi-modal Agents become popular, more such tools optimized for LLM workloads will emerge.