aiXcoder open-sources aiXcompact-4B context compression model

On November 11, 2026, aiXcoder, an AI coding startup incubated from the Institute of Software Engineering at Peking University, open-sourced aiXcompact-4B, a dedicated context compression model. According to a zhidongxi report that day, the model has 4B parameters, targets long-horizon complex development tasks handled by coding agents, and can be deployed privately on a single A100-40GB GPU.

During context compression, aiXcompact-4B structures the coding agent's historical trajectory: it preserves key raw information, trims tool returns such as file reads, code searches and test runs, drops duplicate or no-longer-useful intermediate records, and produces a structured task summary. The aim is to cut token overhead from repeated context re-entry at a lower cost.

Figures released by aiXcoder put the model's measured performance as follows. In the Terminal-Bench benchmark, when aiXcompact-4B was plugged into OpenCode to handle context compression, overall task completion accuracy was on par with DeepSeek-V4-Pro. In private deployment, per-task single GPU inference cost dropped 78% versus DeepSeek-V4-Pro. When called via API, based on published API pricing estimates, one context compression request costs roughly 0.019 CNY in token charges, about 97% below the API cost of DeepSeek-V4-Pro. What is currently verifiable is limited to the test data and cost estimates disclosed by aiXcoder.