Core Insights - Zhipu AI has launched and open-sourced GLM-5, with multiple domestic chips completing Day0 adaptation for optimized performance [1][2] - GLM-5 has achieved state-of-the-art (SOTA) performance in coding and agent capabilities, with significant improvements in parameter scale and pre-training data [2] Group 1: GLM-5 Launch and Adaptation - Zhipu AI's GLM-5 has been adapted for high throughput and low latency on domestic chip platforms including Huawei Ascend, Moore Threads, and others [2] - The collaboration with Haiguang Information has optimized the underlying operators and hardware acceleration for stable operation on Haiguang DCU [1] - Moore Threads has completed full-process adaptation and verification on its MTT S5000 GPU, achieving high-performance inference while reducing memory usage [1] Group 2: Technical Enhancements and Performance Metrics - GLM-5's parameter scale has expanded from 355 billion (with 32 billion activated) to 744 billion (with 40 billion activated), and pre-training data has increased from 23 terabytes to 28.5 terabytes [2] - The introduction of the "Slime" asynchronous reinforcement learning framework supports larger model scales and complex tasks, enhancing learning from long-term interactions [2] - The integration of DeepSeek Sparse Attention mechanism significantly reduces deployment costs while maintaining long text performance [2] Group 3: Support from Huawei - Huawei's Ascend NPU and MindSpore AI framework provide comprehensive support from data to training, contributing to the development of SOTA models based on self-innovated computing power [3]
多款国产芯片宣布Day0支持智谱GLM-5