---
license: apache-2.0
title: README
sdk: static
emoji: 📚
colorFrom: blue
colorTo: purple
---
<div align="center">

<br>

<img src="https://img.shields.io/badge/%E2%9C%A6-TSUKI-000000?style=for-the-badge&labelColor=000000" alt="Tsuki" height="50">

<br><br>

# Token compression for LLMs.

**Feed verbose. Get minimum.**<br>
**Same meaning. Fewer tokens. Lower costs.**

<br>

[![License](https://img.shields.io/badge/Apache_2.0-222222?style=flat-square&logo=apache&logoColor=white)](#)
&nbsp;
[![Flan-T5](https://img.shields.io/badge/Flan--T5_Base-222222?style=flat-square&logo=google&logoColor=white)](#)
&nbsp;
[![Bilingual](https://img.shields.io/badge/ES%20%7C%20EN-222222?style=flat-square&logoColor=white)](#)
&nbsp;
[![15K pairs](https://img.shields.io/badge/15K_training_pairs-222222?style=flat-square&logoColor=white)](#)

<br>

</div>

---

<br>

**Tsuki** is a token compression model. It takes verbose prompts, markdown, and technical instructions and returns the semantic minimum — stripping pleasantries, redundant context, and filler without touching critical information.

Built to reduce costs on paid LLM APIs. Fine-tuned on 15,000 real bilingual compression pairs. Knows when to compress and, more importantly, when not to.

<br>

---

<br>

<div align="center">

```
verbose input → Tsuki → minimum tokens → GPT / Claude / any LLM API
```

</div>

<br>

---

<br>

<div align="center">

**Built with real prompts and zero surrealist dragons.**

<br>

[![Tsuki](https://img.shields.io/badge/Tsuki-2026-000000?style=for-the-badge)](https://huggingface.co/tsuki-team)

</div>