Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
The first open-source compression model trained specifically for coding agents. 4B, Apache 2.0.
| Date | Stars |
|---|---|
| 2026-07-31 | 342 |
| 2026-08-04 | 344 |
| 2026-08-06 | 344 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
<h1 align="center">Paritok</h1>
<p align="center"><b>The first open-source compression model trained specifically for coding agents.</b></p>
<p align="center">Trained on <b>45K real coding-agent trajectories</b>, Paritok understands the difference between a function signature and a debug line — so it keeps what matters and drops what doesn't. Fully compatible with <b>Claude Code, Cursor, OpenHands</b>, and any agent framework using standard message format.<br/><br/><b>~74% fewer tokens on typical workloads</b> (up to <b>95%</b> on heavy long-session traffic), cutting your input token bill by <b>up to 95%</b> on Claude / GPT — while <b>matching gpt-4.1-mini</b> on SWE-bench Verified at a fraction of the cost — deployed as a <b>non-destructive gateway</b> where nothing is ever permanently discarded.</p>
<p align="center">
<a href="https://huggingface.co/paritok/paritok-4b-v1">
<img src="https://img.shields.io/badge/🤗%20Model-HuggingFace-yellow" alt="HF Model"/>
</a>
<a href="./LICENSE">
<img src="https://img.shields.io/badge/License-Apache_2.0-blue" alt="License"/>
</a>
<img src="https://img.shields.io/badge/backbone-Qwen3--4B-purple" alt="Qwen3-4B"/>
<img src="https://img.shields.io/badge/python-3.11+-blue" alt="Python"/>
</p>
<p align="center">
<a href="#-highlights">Highlights</a> ·
<a href="#-benchmark-swe-bench-verified">Benchmark</a> ·
<a href="#-cost-impact">Cost</a> ·
<a href="#-how-paritok-compares">Compare</a> ·
<a href="#-quick-start">Quick Start</a> ·
<a href="#-model-card">Model Card</a> ·
<a href="#-training">Training</a> ·
<a href="#-team">Team</a> ·
<a href="#-citation">Citation</a>
</p>
---
## 📢 News
- **2026-07-14** **Paritok-4B-v1** released on Hugging Face Hub with full SWE-bench Verified end-to-end evaluation.
- **2026-06-25** Finished training. 45K teacher-distilled samples on the Qwen3-4B backbone.
---
## ✨ Highlights
- 🎨 **Code-native.** Trained end-to-end on real coding-agent trajectories (`file_read`, `bash_command`, `log_output`, ...). Paritok knows what an import statement is worth vs a debug line, so it protects function names, paths, and error strings while compressing.
- 🚀 **Up to 95% fewer tokens** (74% on typical workloads) — compresses each segment to **25.7%** of original; drop-aware long sessions push overall cuts to 95%. **2× harder than gpt-4.1-mini** (50.2% CR) and **2.4× harder than gpt-5** (61.9% CR).
- 🎯 **Retains 86.5% of full-context solve quality** on SWE-bench Verified — matching gpt-4.1-mini as compressor at **less than half the token spend**, and within 7pp of gpt-5 at **40% its context length**.
- 💰 **Up to 95% off your input token bill** (74% on typical workloads) at Claude Sonnet pricing. Long-session teams save **thousands per month** — see [Cost Impact](#-cost-impact).
- 🪶 **Small & self-hostable** — 4B LoRA adapter, bf16, runs on a single 24GB GPU. No SaaS, no lock-in, no per-token compressor fee.
- 🔁 **Non-destructive by design.** Compressed content is never gone — the agent recalls any exact original on demand via `expand_context`. Lossy on the wire, fully recoverable when it counts. The 86.5% benchmark is the raw model *without* recall — a conservative floor, not the ceiling you actually run.
- 🔓 **Fully open** — Apache 2.0 weights, reproducible data pipeline, real end-to-end SWE-bench numbers (no cherry-picking).
---
## 📊 Benchmark: SWE-bench Verified
Real end-to-end evaluation on **SWE-bench Verified**. An agent scaffold receives its context through each compressor, then attempts to resolve the issue. Primary metric is **quality retained** (solve rate normalized to the uncompressed baseline) — the fair way to compare compressors of different aggressiveness.
### Solve quality vs compression rate
| Context source | **Quality retained** ¹ | Compression rate |
| ------------------------- | :--------------------: | :--------------: |
| Uncompressed baseline | 100.0% | Excerpt of 28,898 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:b8b6778b9f9ce2c0, topic:developer-tools