PACKLM.

Back to Works
WorksAI tooling / Open Source • 2026

PackLM

The 80% Smaller JSON Alternative for LLMs

PackLM
RoleCreator / Full Stack Developer
Year2026
Client / OriginIndependent / Open Source
CategoryAI tooling / Open Source

About the project

PackLM is an innovative data serialization format specifically designed for Large Language Models. By separating the schema from the actual data into a plain text format, it reduces token usage by up to 80% compared to standard JSON when sending structured data to AI models, significantly cutting down API costs and latency.

The Challenge

When building applications powered by Large Language Models (LLMs), developers typically rely on JSON to send structured data to the model.

However, JSON is notoriously verbose. Every single object repeats its keys (e.g., `{"name": "Alice", "age": 25}, {"name": "Bob", "age": 30}`). When sending arrays with hundreds or thousands of items to an LLM, this repetition consumes a massive amount of tokens.

This leads to two major problems:

  1. High Costs: You pay for every token sent to the API.
  2. Context Limits & Latency: Verbose JSON eats up the context window quickly and slows down the time-to-first-token.

The Solution

I built PackLM to solve the token bloat issue. It works by completely decoupling the data schema from the actual values.

Instead of repeating keys, PackLM defines the structure once at the top, and then lists the values in a highly compressed, CSV-like plain text format that LLMs naturally understand.

Key Achievements:

  • Achieved up to 80% reduction in token usage compared to standard minified JSON.
  • Built a lightweight parser/serializer that seamlessly integrates into existing AI pipelines.
  • Designed a modern, interactive landing page to demonstrate the token savings in real-time.

No gallery images available

Next Project
CodeShip