Codegeex
CodeGeeX is a large-scale multilingual code generation model developed by Tsinghua University that excels at cross-language code translation, trained on 20+ programming languages with official IDE plugins for VS Code and JetBrains — representing China's leading contribution to the open-source code generation ecosystem and demonstrating that carefully curated multilingual training data produces models with superior code translation capabilities compared to English-centric alternatives.
---
Architecture & Training
| Component | Detail |
|---|---|
| Parameters | 13B (CodeGeeX-1), upgraded in CodeGeeX-2 |
| Architecture | Decoder-only transformer with custom positional encodings |
| Training Data | 850GB of code across 23 programming languages |
| Hardware | Trained on Ascend 910 AI processors (Huawei) — not NVIDIA GPUs |
| Languages | Python, C++, Java, JavaScript, Go, Rust, and 17 others |
| Context | 2048 tokens |
A notable technical distinction: CodeGeeX was trained on Huawei Ascend hardware, demonstrating that competitive AI models can be built on non-NVIDIA infrastructure — strategically important given US export restrictions on advanced chips to China.
---
Cross-Language Translation
CodeGeeX's standout capability is code translation between programming languages:
The model can translate functions between any pair of its 23 supported languages with high accuracy — converting Python data processing scripts to optimized C++ implementations, translating Java enterprise code to modern Kotlin, or porting JavaScript web applications to TypeScript with proper type annotations.
This capability emerges from the balanced multilingual training — unlike English-centric models that treat non-Python languages as secondary, CodeGeeX allocates proportional training compute to each language family, producing more uniform cross-language competence.
---
🏗️ IDE Integration & CodeGeeX2
IDE Plugins: CodeGeeX provides polished, production-grade extensions for VS Code and JetBrains IDEs (IntelliJ, PyCharm, WebStorm) with features including inline completion, code explanation, comment-to-code generation, and the signature cross-language translation tool.
CodeGeeX2 (based on ChatGLM2-6B) dramatically improved on the original:
- 6B parameters (smaller but more efficient than the 13B v1)
- 10x faster inference through quantization and architecture optimization
- Support for repository-level context understanding
- Chat-based coding assistance (explain, debug, refactor)
CodeGeeX has over 500,000 active users across its IDE plugins, making it one of the most widely deployed open-source code assistants globally — particularly dominant in the Chinese developer ecosystem.