GLM-4-Extreme
About
GLM-4-Extreme, while not specifically documented, seems to be a variant under the GLM-4 model family developed by Zhipu AI. The GLM-4 models are noted for their strong performance on various benchmarks, including MMLU and HumanEval, as well as their proficiency in instruction following in multiple languages, primarily English and Chinese. They support a long context length of up to 128k tokens, with some versions even reaching 1M tokens. Additionally, these models feature integrated tool usage for complex tasks and voice capabilities for real-time speech processing. The GLM-4 architecture is based on a multi-stage training process, combining pre-training, supervised fine-tuning, and reinforcement learning with human feedback. However, the existence of the GLM-4-Extreme variant specifically lacks documentation, suggesting it may not be officially detailed or released.