Hy4 previewvsGLM-5.3
在 12 个共同 benchmark 中，GLM-5.3 整体领先：Hy4 preview 领先 2 项，GLM-5.3 领先 10 项，持平 0 项，平均分差 -12.14。
Hy4 preview
腾讯AI实验室 · 2026-08-28 · 推理大模型
GLM-5.3
智谱AI · 2026-08-14 · 推理大模型
Hy4 preview2 项(17%)(83%)10 项GLM-5.3
评测分数
按能力类目分组，每组内按分差大小排列；共 12 项。
Coding and Software Engineer
GLM-5.3 领先 4/5| 评测项 | Hy4 preview | GLM-5.3 | 分差 |
|---|---|---|---|
| SWE-Marathon | 31.903 / 5Thinking High (With Tools) | 42.501 / 5Max (With Tools) | -10.60 |
| PostTrain Bench | 35.604 / 5Thinking High (With Tools) | 39.801 / 5Max (With Tools) | -4.20 |
| DeepSWE | 64.3011 / 31Thinking High (With Tools) | 66.908 / 31Max (With Tools) | -2.60 |
| Program Bench | 17.506 / 6Thinking High (With Tools) | 195 / 6Max (With Tools) | -1.50 |
| NL2Repo-Bench | 58.902 / 11Thinking High (With Tools) | 583 / 11Max (With Tools) | +0.90 |
AI Agent - Tool Usage
GLM-5.3 领先 3/4| 评测项 | Hy4 preview | GLM-5.3 | 分差 |
|---|---|---|---|
| AutomationBench | 32.103 / 10Thinking High (With Tools) | 48.202 / 10Max (With Tools) | -16.10 |
| CyberGym | 78.403 / 4Thinking High (With Tools) | 84.501 / 4Max (With Tools) | -6.10 |
| Terminal-Bench 2.1 | 85.4010 / 47Thinking High (With Tools) | 88.203 / 47Max (With Tools) | -2.80 |
| Toolathlon-Verified | 74.103 / 8Thinking High (With Tools) | 736 / 8Max (With Tools) | +1.10 |
Agent Level Benchmark
GLM-5.3 领先 1/1| 评测项 | Hy4 preview | GLM-5.3 | 分差 |
|---|---|---|---|
| Agents' Last Exam | 22.8013 / 14Thinking High (With Tools) | 28.504 / 14Max (With Tools) | -5.70 |
General Knowledge
GLM-5.3 领先 1/1| 评测项 | Hy4 preview | GLM-5.3 | 分差 |
|---|---|---|---|
| HLE | 55.4015 / 185Thinking High (With Tools) | 62.503 / 185Max (With Tools) | -7.10 |
Productivity Knowledge
GLM-5.3 领先 1/1| 评测项 | Hy4 preview | GLM-5.3 | 分差 |
|---|---|---|---|
| GDPval-AA v2 | 1,6788 / 16Thinking High (With Tools) | 1,7693 / 16Max (With Tools) | -91 |
规格对比
| 字段 | Hy4 preview | GLM-5.3 |
|---|---|---|
| 发布机构 | 腾讯AI实验室 | 智谱AI |
| 发布时间 | 2026-08-28 | 2026-08-14 |
| 模型类型 | 推理大模型 | 推理大模型 |
| 架构 | MoE 架构 | MoE 架构 |
| 参数规模 | 7700亿 | 7440亿 |
| 上下文长度 | 1M | 1M |
| 最大输出 | 暂无数据 | 128K |
API 调用价格
价格优先使用 DataLearner 配置的 API 记录；缺失项不做推测。
| 价格项 | Hy4 preview | GLM-5.3 |
|---|---|---|
| 文本输入 | ¥6 / 1M tokens | 暂无公开价格 |
| 文本输出 | ¥18 / 1M tokens | 暂无公开价格 |
| 缓存读取 | ¥0.3 / 1M tokens | 暂无公开价格 |
部分模型公开价格不完整，缺失字段按"暂无公开价格"展示。
小结
- GLM-5.3在以下类目领先:Coding and Software Engineer (4/5)、AI Agent - Tool Usage (3/4)、Agent Level Benchmark (1/1)、General Knowledge (1/1)、Productivity Knowledge (1/1)
12 个共同 benchmark 上，GLM-5.3 平均高出 12.14 分。
单项差距最大的 benchmark：GDPval-AA v2 — Hy4 preview 1,678，GLM-5.3 1,769（分差 -91）。
本页正文由结构化模型、价格与 benchmark 数据生成，不使用实时 LLM 撰写。