codegeex-4

CodeGeeX4-ALL-9B is an open-source, professional code-generation model developed by the Tsinghua University team, with a core focus on being a multilingual, all-purpose development assistant that supports ultra-long contexts of up to 128K tokens. It is continuously trained based on GLM-4-9B and achieves enterprise-level performance with lightweight parameters, covering all development scenarios—including code completion and repository-level question answering.

• Performance Breakthrough: On authoritative benchmarks such as BigCodeBench and NaturalCodeBench, it sets a new record for 10B-parameter models with a score of 48.9 points, surpassing large models like Llama3-70B
• Scenario Adaptation: It uniquely supports function calls, code interpreters, and repository-level code question answering, covering the entire development workflow from single-line completion to project refactoring
• Efficiency Advantage: Its inference cost is only 1/4 of comparable models, and its development cycle is shortened by 50%, making it especially suitable for resource-sensitive teams
• Long-Text Processing: With support for 128K contexts, its code retrieval accuracy reaches 100% (CRUXEval test)
• Well-Established Ecosystem: It provides VS Code/Jetbrains plugins, local deployment guides, and multi-framework deployment solutions (vLLM/Candle)

───────────────────────────────────────────────────────────────────

Core Capabilities

⚡ Ultra-Long Context Handling: With a sequence length of 128K tokens, it easily handles project analysis involving millions of lines, enabling precise code retrieval and cross-file completion

🌐 Multilingual Code Generation: It supports over 73 languages including Python and Rust, achieving a score of 82.3 on the HumanEval test—surpassing DeepSeek Coder 33B

🔧 Enterprise-Level Function Calls: It uniquely implements Function Call capability, with a success rate exceeding GPT-4 and directly interfacing with APIs and database operations

📊 Intelligent Repository-Level Interaction: It completes code review and refactoring suggestions in a single model, validated by NaturalCodeBench with a score of 40.4, and supports cross-file understanding

🚀 Rapid Deployment Solution: It is compatible with vLLM/Ollama/Candle frameworks, providing out-of-the-box OpenAI-compatible interfaces and allowing you to set up a local service in just 5 minutes