Paritok-4B compresses coding-agent context to 25.7% of size with a 264 MB adapter, and shows gpt-5 as a compressor is net-negative

arXiv·high signalXBlueskyLinkedInCopy link

Released 25 August (arXiv 2608.24188), Paritok-4B is a 4B LoRA on Qwen3-4B distilled from a gpt-4.1-mini teacher over 67,074 real OpenHands trajectories into 40,606 validated examples. It is extractive rather than paraphrasing (96.0% of emitted identifiers, paths, and numbers already appear in its input) and intent-conditioned on the agent's current task, compressing to 25.7% of original size on all 300 SWE-bench Lite instances while retaining 86.5% of uncompressed single-shot solve quality, versus 50.2% for a gpt-4.1-mini compressor and 61.9% for gpt-5. The economics are the point: it self-hosts on one 24 GB GPU with no per-token compressor fee, and at list prices gpt-5 as a compressor costs more than the downstream tokens it saves. Weights, data, and eval scripts are Apache 2.0.