Inference cost reduced by 78%, API cost reduced by 97%, Peking University-affiliated startup open-sources 4B context compression model

What happened: aiXcoder, a startup incubated from Peking University's software engineering institute, open-sourced aiXcompact-4B, a 4B context compression model for coding agents, deployable on a single A100-40GB GPU. Key numbers: Inference cost down 78% versus DeepSeek-V4-Pro in private deployment. Why it matters: aiXcoder positions the model as a lightweight, economical option for high-frequency context compression, part of a large-model plus small-model split aimed at lowering enterprise compute costs for long agent tasks.

aiXcoder, a startup incubated from Peking University's software engineering institute, open-sourced aiXcompact-4B, a 4B context compression model for coding agents, deployable on a single A100-40GB GPU.

aiXcoder positions the model as a lightweight, economical option for high-frequency context compression, part of a large-model plus small-model split aimed at lowering enterprise compute costs for long agent tasks.