GLM-4 Air

GLM-4 Air Model Introduction

Key Capabilities and Primary Use Cases

• Multilingual Support: Primarily aligned for Chinese and English, with additional support for 24 languages.
• Task Completion: Capable of accessing online information via web browsing, using Python interpreters for math problems, leveraging text-to-image models, and calling user-defined functions[2][5].
• Instruction Following: Effective in following instructions in both English and Chinese[2].

Most Important Features and Improvements

• Multi-Stage Post-Training: Includes supervised fine-tuning and learning from human feedback to achieve high-quality alignment[2].
• Efficient Inference: Designed for lower latency and inference cost compared to GLM-4, making it more efficient for real-time applications[2][3].
• Advanced Tools Integration: Automatically selects and uses appropriate tools to complete complex tasks[2].

Essential Technical Specifications

• Training Data: Pre-trained on ten trillions of tokens, mostly in Chinese and English[2].
• Context Length: Supports up to 128K context length[2][5].
• Model Versions: Available in various versions, including GLM-4-Air, GLM-4-9B, and GLM-4V-9B[2][5].

Notable Performance Characteristics

• Benchmark Performance: Closely rivals or outperforms GPT-4 in several benchmarks such as MMLU, GSM8K, MATH, and HumanEval. Outperforms GPT-4 in Chinese alignments and matches GPT-4-Turbo in instruction following[2].
• Long-Context Tasks: Matches the performance of GPT-4 Turbo and Claude 3 in long-context tasks[2].