Date: 2025-12-01 Version: v1.1.0 Environment: Node.js (macOS)
We compared ZON against JSON (native) and MsgPack (msgpack-lite) across three datasets:
- Hiking (Small, Mixed): A typical LLM context object with metadata and a small table.
- Large Array (1k items): A tabular dataset with 1000 rows, testing table compression.
- Nested Object (Deep): A deeply nested structure to test recursion overhead.
| Format | Size (bytes) | Tokens (GPT) | Encode (ms) | Decode (ms) |
|---|---|---|---|---|
| JSON | 366 | 115 | 0.035 | 0.016 |
| MsgPack | 277 | N/A | 0.145 | 0.045 |
| ZON | 278 | 96 | 0.180 | 0.120 |
Result: ZON saves 16.5% tokens vs JSON.
| Format | Size (bytes) | Tokens (GPT) | Encode (ms) | Decode (ms) |
|---|---|---|---|---|
| JSON | 97,891 | 31,002 | 1.520 | 0.850 |
| MsgPack | 88,903 | N/A | 3.200 | 1.800 |
| ZON | 68,904 | 25,004 | 15.400 | 8.500 |
Result: ZON saves 19.3% tokens vs JSON.
| Format | Size (bytes) | Tokens (GPT) | Encode (ms) | Decode (ms) |
|---|---|---|---|---|
| JSON | 4,012 | 1,480 | 0.014 | 0.033 |
| MsgPack | 2,689 | N/A | 0.199 | 0.124 |
| ZON | 3,330 | 1,366 | 0.221 | 0.916 |
Result: ZON saves 7.7% tokens vs JSON.
- Token Efficiency: ZON consistently outperforms JSON in token count, with savings ranging from 7% to 19%. This directly translates to lower LLM costs and larger effective context windows.
- Size: ZON is significantly smaller than JSON in bytes, often approaching MsgPack's binary size for tabular data due to header deduplication.
- Performance: While slower than native JSON (as expected), ZON's performance is well within acceptable limits for real-world applications, processing 1000 items in ~15ms.
Benchmarks were run using benchmarks/performance.ts.
- JSON:
JSON.stringify/JSON.parse - MsgPack:
msgpack-lite - ZON:
zon-formatv1.1.0 - Tokens: Measured using
gpt-tokenizer(GPT-3.5/4 encoding).