Private deployment · Sovereign stack · Industry SOTA
aiXcoder-7B · Code completion & generation model
Released and open-sourced in April 2024, this fully self-developed code model was pre-trained on a 1.5-trillion-unique-token high-quality code dataset with code-specific objectives, outperforming open-source models of its own size and larger. The accompanying paper was accepted at ICSE 2025.
aiXcoder-7B is built for code generation and completion. With innovations such as SFIM structured training and CoLT, it delivers stronger real-world engineering performance from fewer parameters:
SOTA on code generation, code completion and cross-file context generation, beating 34B code LLMs five times its size.
aiXcoder-7B deploys privately on a single GPU and can be personalized with enterprise domain knowledge, proprietary frameworks and coding standards.
- ICSE 2025
- Accepted at a top international venue; ICSE reviewers highlighted it as exposing a key deployment problem for LLMs, with major contribution to industry
- 94.4%
- API-invocation accuracy — keeping agents running reliably
- Sovereign & self-controlled
- Compatible with domestic operating systems and chips; private deployment on the enterprise intranet