qwen-coder-plus-2024-11-06

The entire Qwen-coder-plus series is an open-source family of code models launched by Alibaba Tongyi, characterized by being “powerful, diverse, and practical.” Its core positioning is as a “highly efficient programming assistant covering all model sizes,” suitable for a wide range of scenarios—from lightweight edge-side development to complex enterprise-level projects, with code-generation capabilities reaching the top level in the open-source community.

• Industry-Leading Performance (Open-Source SOTA): The flagship 32B-Instruct model matches GPT-4o in performance, ranking first among open-source models on key benchmarks such as HumanEval (92.7 points) and MBPP (90.2 points). It excels in code repair and reasoning tasks, achieving a 68.9% win rate against GPT-4 Turbo.
• Deep Multilingual Coverage: Supports over 40 programming languages, topping open-source rankings in both McEval (65.9 points) and MdEval (75.2 points). It demonstrates strong capabilities in handling niche languages, significantly reducing the learning costs associated with unfamiliar programming languages.
• Diverse Model Sizes: Covers six parameter scales from 0.5B to 32B. Each size offers both Base (fine-tuned foundation) and Instruct (ready-to-use) versions. Model performance improves positively with increasing parameters, making it adaptable to various hardware resources.
• Practical Scenario Adaptation: Integrates seamlessly with tools like Cursor; the 32B-Base model achieves SOTA results in five key benchmarks for code completion. It can also generate visualized outputs and will soon introduce a “one-sentence app generation” feature.

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

Core Capabilities

💻 High-Precision Code Generation: Generates runnable code, supporting algorithm implementation, full-stack development, and demonstrating outstanding out-of-distribution (OOD) generalization on new problem sets.

🔧 Code Repair and Optimization: Accurately identifies vulnerabilities and leads the open-source community in multilingual repair capabilities, dramatically boosting programming efficiency.

🧠 Strong Code Reasoning: Understands code execution logic and accurately predicts input-output behavior, greatly facilitating complex development tasks.

📚 Long-Text Code Understanding: Supports up to 128K context for models of 7B parameters and above, making it ideal for large-scale codebases and cross-file analysis.

🌍 Multi-Scenario Compatibility: Suitable for novice programmers, professional developers, and enterprise teams alike, with friendly Chinese language support.