Paritok 4B

Paritok optimizes LLM usage for software developers by intercepting requests from coding agents to LLM APIs and compressing input context before sending upstream. This middleware approach achieves approximately 74% token reduction on typical workloads, with up to 95% savings on extended sessions.

Highlights

• Achieves 74-95% token compression while retaining 86.5% solve quality on SWE-bench Verified
• Trained on 45,000 real coding-agent trajectories to understand developer workflows
• Non-destructive compression with built-in `expand_context` tool for recovery if needed
• Code-aware design that filters 70+ tool schemas using semantic embeddings
• Self-hostable 4B LoRA adapter under Apache 2.0 license

Perfect for developers running coding agents like Claude Code, Cursor, or OpenHands who want to lower API costs and latency. Paritok runs free on your own 24GB GPU or offers optional per-request billing through their hosted service. No subscription required.