The ZhipuAI Open Platform recently announced that the GLM-4-Flash model is now available to the public for free. Users can register at bigmodel.cn to build their own models and applications at no cost, a move expected to trigger a new wave of innovation in the AI application field.
Known for its speed and cost-effectiveness, the GLM-4-Flash model is particularly suitable for simple tasks requiring quick responses. It supports multi-round conversations, web browsing, Function Call, and long-text reasoning, covering 26 languages including Chinese, English, Japanese, Korean, and German. Its applications are diverse, ranging from scientific data preprocessing and information extraction to multilingual translation and multi-round dialogues.
Technically, GLM-4-Flash optimizes its efficiency through adaptive weight quantization, parallelization, batch processing strategies, and speculative sampling, significantly reducing inference costs. During pre-training, it utilized large language models to filter data, acquiring over 10T of high-quality multilingual data, more than three times that of the ChatGLM3-6B model. Additionally, the application of FP8 technology greatly enhanced pre-training efficiency.
In terms of performance, GLM-4-Flash excels in logical reasoning tests, with a generation speed of 72.14 tokens/s (approximately 115 characters/s), second only to GPT-3.5-turbo. Furthermore, the model possesses real-time web retrieval capabilities, allowing quick access to information such as weather and news, and its function calling ability is on par with GPT-4.
The applications of GLM-4-Flash have shown potential in various fields. Biologists use it to process molecular data and explore health codes; developers generate training data to help other large models grow; university scholars create AI teaching assistants to enhance learning experiences; translation app developers use it to develop interactive English learning tools for children; and advertising agencies use it for text polishing to improve copywriting efficiency.
To encourage user feedback, the ZhipuAI Open Platform is offering a limited-time free fine-tuning benefit: complete the survey to receive a 5 million Tokens (3 months) training resource package, available on a first-come, first-served basis.
Experience URL: https://bigmodel.cn/console/trialcenter?modelCode=glm-4-flash