9月30日，国产大模型"六小龙"之一的智谱发布GLM-4.6新模型。 作为GLM系列最新版本，GLM-4.6在真实编程、长上下文处理、推理能力、信息搜索、写作能力与智能体应用等多个方面能力有所提升。 官方信息显示，此次升级表现在公开基准与真实编程任务中，GLM-4.6代码能力对齐Claude Sonnet4；上下文窗口由128K提升至200K，适应更长的代码和智 能体任务；新模型提升推理能力，并支持在推理过程中调用工具；搜索方面增强模型的工具调用和搜索智能体。 另外，"模芯联动"是此次新模型发布的重点，GLM-4.6已在寒武纪国产芯片上实现FP8+Int4混合量化部署，这也是行业首次在国产芯片上投产的FP8+Int4模 型芯片一体解决方案，在保持精度不变的前提下，降低推理成本，为国产芯片在大模型本地化运行上探索可行路径。 具体到模型适配过程中，占总内存的60%-80%的大模型核心参数通过Int4量化后，可将权重体积直接压缩为FP16的1/4，大幅降低芯片显存的占用压力；推理 环节积累的临时对话数据可以通过Int4压缩内存的同时，将精度损失控制在"轻微"范围。而FP8可重点针对模型中"数值敏感、影响推理准确性" ...

官方信息显示，此次升级表现在公开基准与真实编程任务中，GLM-4.6代码能力对齐Claude Sonnet 4； 上下文窗口由128K提升至200K，适应更长的代码和智能体任务；新模型提升推理能力，并支持在推理 过程中调用工具；搜索方面增强模型的工具调用和搜索智能体。 另外，"模芯联动"是此次新模型发布的重点，GLM-4.6已在寒武纪国产芯片上实现FP8+Int4混合量化部 署，这也是行业首次在国产芯片上投产的FP8+Int4模型芯片一体解决方案，在保持精度不变的前提下， 降低推理成本，为国产芯片在大模型本地化运行上探索可行路径。 FP8是8位浮点数（Floating-Point 8）数据类型，动态范围广、精度损失小；Int4是4 位整数（Integer 4） 数据类型，压缩比极高，内存占用最少，适配低算力硬件但精度损失相对明显。此次尝试的"FP8+Int4 混合" 模式，并非简单将两种格式叠加，而是根据大模型的"模块功能差异"，针对性分配量化格式，让 该省内存的地方用Int4压到极致，该保精度的地方用FP8守住底线，实现合理资源分配。 具体到模型适配过程中，占总内存的60%-80%的大模型核心参数通过Int ...

Core Insights - The latest model GLM-4.6 from the domestic AI startup Zhipu has been released, showcasing improvements in programming, long context handling, reasoning capabilities, information retrieval, writing skills, and agent applications [3] Model Enhancements - GLM-4.6 aligns its coding capabilities with Claude Sonnet 4 in public benchmarks and real programming tasks [3] - The context window has been increased from 128K to 200K, allowing for longer code and agent tasks [3] - The new model enhances reasoning abilities and supports tool invocation during reasoning processes [3] - There is an improvement in the model's tool invocation and search capabilities [3] Chip Integration - The "MoCore linkage" is a key focus of the new model, with GLM-4.6 achieving FP8+Int4 mixed quantization deployment on domestic Cambricon chips, marking the industry's first production of an FP8+Int4 model chip solution on domestic hardware [3] - This approach maintains accuracy while reducing inference costs, exploring feasible paths for localized operation of large models on domestic chips [3] Quantization Techniques - FP8 (Floating-Point 8) offers a wide dynamic range with minimal precision loss, while Int4 (Integer 4) provides high compression ratios with lower memory usage but more noticeable precision loss [4] - The "FP8+Int4 mixed" mode allocates quantization formats based on the functional differences of the model's modules, optimizing memory usage [4] Memory Efficiency - Core parameters of the large model, which account for 60%-80% of total memory, can be compressed to 1/4 of FP16 size through Int4 quantization, significantly reducing the memory pressure on chips [5] - Temporary dialogue data accumulated during inference can also be compressed using Int4 while keeping precision loss minimal [5] - FP8 is used for numerically sensitive modules to minimize precision loss and retain fine semantic information [5] Ecosystem Development - Cambricon and Moore Threads have successfully adapted GLM-4.6 based on the vLLM inference framework, demonstrating the capabilities of the new generation of GPUs to run the model stably at native FP8 precision [5] - This adaptation signifies that domestic GPUs are now capable of collaborating and iterating with cutting-edge large models, accelerating the development of a self-controlled AI technology ecosystem [5] - The combination of GLM-4.6 and domestic chips will be offered to enterprises and the public through the Zhipu MaaS platform [5]