Paritok
Paritok provides a non-destructive compression gateway for AI coding agents to reduce input-token bills by up to 85% without changing agent logic.
At a Glance
- Developers using AI coding tools
- AI startups managing token costs
- Engineering organizations deploying coding agents
AI Tools by Paritok
(1)
Paritok
Token Compression Proxy for LLMs
Discussions
No discussions yet
Be the first to start a discussion about Paritok
Latest News
Conclusion of the Build with Paritok: The Token-Efficiency Hackathon
Paritok v1.3.0 Release: Stability and API Improvements
Paritok v1.2.0 Release: Embedding-based Tool Selection Ships
Paritok Gateway Open-Sourced
Products & Services
Open-source compression model trained on coding trajectories, shrinking input segments to ~26% of original size.
A non-destructive middleware proxy that rewrites agent requests to compress tool schemas, file reads, and history.
Managed endpoint for compression, provided for free during launch and integrated with a usage dashboard.
Market Position
The first open-source 4B model trained end-to-end on real coding-agent trajectories, beating LLMLingua-2 and gpt-4.1-mini in compression and quality.
Leadership
Founders
Jiayu Shi
Co-founder of Paritok. Responsible for training, modeling, reward design, and the data pipeline. Lead judge for the 'Build with Paritok' hackathon.
Luzhuo Chen
Co-founder of Paritok. Responsible for evaluation, deployment, product management, and the data pipeline. Lead judge for the 'Build with Paritok' hackathon.
Executive Team
Jiayu Shi
Co-founder
Training and modeling expert for Paritok.
Luzhuo Chen
Co-founder
Evaluation and deployment lead for Paritok.
Founding Story
Built by two engineers (Jiayu Shi and Luzhuo Chen) who invested months of GPU budget to create the first open-source compression model specifically for coding trajectories.
Business Model
Revenue Model
Free self-hosting (Apache 2.0); usage-based pricing for hosted GPU services.
Pricing Tiers
Apache-2.0 open-source version; runs on local hardware via Ollama or vLLM.
Managed server; free until August 31, 2026.
Target Markets
- Developers using AI coding tools
- AI startups managing token costs
- Engineering organizations deploying coding agents
- AI-assisted coding with Claude Code/Cursor/Codex
- Context-heavy codebase auditing
- Reducing API costs for long agent sessions