GLM-5.3-Flash
A lightweight large language model in Zhipu AI's GLM series, supporting multi-turn dialogue and tool calling, with optimization targeting inference speed.
Core Facts
Timeline (last 90 days)
GLM-5.3 has an input multiplier of 6.9, a cached input multiplier of 1.7, and an output multiplier of 24; GLM-5.3-Flash has an input multiplier of 2.3, a cached input multiplier of 0.56, and an output multiplier of 8.
GLM-Legal's score on Harvey LAB is close to three times that of its base model GLM-5.3-Flash.
GLM-Legal is built on the base model GLM-5.3-Flash, belonging to the 'base + domain fine-tuning' route.
几乎所有对比数据都来自 Z.ai 自报,且各基准的测试框架设置不同,跨模型差距应视为方向性参考而非定论。
Artificial Analysis 测得 GLM-5.3-Flash 输出速度为48.7 tokens/秒,首token时间1.52秒,智能指数57分(前沿级别)。
GLM-5.3-Flash 在 DeepSWE v1.1 得63.4分,比 GLM-5.2 的46.2分提升17.2分;HLE 得55.3分。
GLM-5.3-Flash 在 Terminal-Bench 2.1 得84.3分,略低于 Claude Opus 4.8 的85.0分和 GPT-5.6 Terra 的87.4分。
GLM-5.3-Flash 在 Toolathlon Verified 得78.4分,OfficeQA Pro 得62.4分,均领先 Claude Opus 4.8。
GLM-5.3-Flash 在国产AI芯片上规模化运行,搭配基于 SGLang 的专用推理引擎,声称端到端服务增益达到3倍。
Z.ai 声称 GLM-5.3-Flash 的注意力计算量减少3.0倍,KV缓存缩小4.4倍,从而支持1M token上下文窗口(提示加输出不超过1,048,576)。
40 more timeline events
All Facts (20)
The Flash version of the GLM series is positioned as a lightweight model with low latency and low cost.
75%VerifiedZhipu AI has launched the GLM-5.3 Flash lightweight model.
85%UnverifiedGLM-5.3-Flash achieved a score of 57 on the Artificial Analysis composite intelligence index evaluation, on par with Claude Opus 4.8.
70%UnverifiedThrough MTP (Multi-Token Prediction) technology, the inference speed of Qwen3.8-Flash-Next and GLM-5.3-Flash is improved by up to 2x.
70%UnverifiedGLM-5.3-Flash, from Z.ai, is a multimodal MoE model with a total of 320B (320 billion) parameters, with only 18B parameters activated at any given time.
70%UnverifiedThe limitation of specialized small models is that their generalization ability rapidly degrades when task boundaries are ambiguous, cross-domain reasoning is required, or unseen edge cases need to be handled.
60%UnverifiedGLM-5.3 has an input multiplier of 6.9, a cached input multiplier of 1.7, and an output multiplier of 24; GLM-5.3-Flash has an input multiplier of 2.3, a cached input multiplier of 0.56, and an output multiplier of 8.
50%UnverifiedGLM-Legal is built on the base model GLM-5.3-Flash, belonging to the 'base + domain fine-tuning' route.
50%UnverifiedGLM-Legal's score on Harvey LAB is close to three times that of its base model GLM-5.3-Flash.
50%UnverifiedGLM-5.3-Flash 在 DeepSWE v1.1 得63.4分,比 GLM-5.2 的46.2分提升17.2分;HLE 得55.3分。
50%UnverifiedGLM-5.3-Flash 的注意力架构将计算拆为线性注意力(O(n)递归状态,处理近期token)和稀疏注意力(轻量索引器回扫全局上下文,只拉取相关远端片段)两条路径。
50%UnverifiedArtificial Analysis 测得 GLM-5.3-Flash 输出速度为48.7 tokens/秒,首token时间1.52秒,智能指数57分(前沿级别)。
50%UnverifiedGLM-5.3-Flash 在国产AI芯片上规模化运行,搭配基于 SGLang 的专用推理引擎,声称端到端服务增益达到3倍。
50%UnverifiedGLM-5.3-Flash 是第一个将稀疏注意力和线性注意力融合进单一混合系统的开放前沿模型。
50%UnverifiedZ.ai 声称 GLM-5.3-Flash 的注意力计算量减少3.0倍,KV缓存缩小4.4倍,从而支持1M token上下文窗口(提示加输出不超过1,048,576)。
50%UnverifiedGLM-5.3-Flash 在 Toolathlon Verified 得78.4分,OfficeQA Pro 得62.4分,均领先 Claude Opus 4.8。
50%UnverifiedGLM-5.3-Flash 在 AutomationBench 上得48.8分,领先 Claude Opus 4.8 的41.0分和 GPT-5.6 Terra 的37.2分,单任务成本约0.045美元。
50%UnverifiedGLM-5.3-Flash 预训练使用了30万亿token的多模态语料,是首个原生多模态的 GLM-5,支持文本、图像、视频和文件。
50%UnverifiedGLM-5.3-Flash 使用 IndexPool 技巧,将索引器的 key 向量按4合1做加权池化,在百万token上下文下保持检索机制小巧。
50%Unverified根据 vLLM 配置和 Z.ai 文档,GLM-5.3-Flash 的45层模型将 KDA 线性注意力层与 NoPE 稀疏 MLA 层交替堆叠。
50%