DeepSeek-V4.1-FlashvsGLM-5.3
在 23 个同模式 benchmark 中，DeepSeek-V4.1-Flash 整体领先：DeepSeek-V4.1-Flash 领先 14 项，GLM-5.3 领先 9 项，持平 0 项，平均分差 -24.51。另有 17 项测试模式不同，仅供参考。
DeepSeek-V4.1-Flash
DeepSeek-AI · 2026-09-10 · 多模态大模型
GLM-5.3
智谱AI · 2026-08-14 · 推理大模型
DeepSeek-V4.1-Flash14 项(61%)(39%)9 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 23 项。
仓库修复与多文件工程
DeepSeek-V4.1-Flash 领先 3/3| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| NL2Repo-Bench | 65.403 / 20Max (With Tools) | 588 / 20Max (With Tools) | +7.40 |
| DeepSWE | 74.208 / 108Max (With Tools) | 68.9632 / 108Max (With Tools) | +5.24 |
| APEX-SWE | 51.9014 / 72Max (With Tools) | 47.5021 / 72Max (With Tools) | +4.40 |
数学
GLM-5.3 领先 2/3| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| Riemann-bench | 2425 / 42Max (With Tools + Internet) | 1234 / 42Max (With Tools + Internet) | +12 |
| FrontierMath Tier 4 v2 | 26.8344 / 77Max (With Tools) | 29.2741 / 77Max (With Tools) | -2.44 |
| FrontierMath v2 | 67.3732 / 119Max (With Tools) | 68.7729 / 119Max (With Tools) | -1.40 |
自主开发与终端任务
GLM-5.3 领先 2/3| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| Terminal-Bench 4.0 | 26.8044 / 110Max (With Tools) | 41.8228 / 110Max (With Tools) | -15.02 |
| Vibe Code Bench 1-100 | 16.3816 / 25Max (With Tools) | 19.9912 / 25Max (With Tools) | -3.61 |
| Terminal-Bench 3.0 | 304 / 11Max (With Tools) | 28.306 / 11Max (With Tools) | +1.70 |
科学计算与科研编程
胶着 2/2| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 15.709 / 19Max (With Tools) | 8.1014 / 19Max (With Tools) | +7.60 |
| SciCode | 51.9063 / 136Max (No Tools) | 599 / 136Max (No Tools) | -7.10 |
跨能力综合测试
DeepSeek-V4.1-Flash 领先 2/2| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| LiveBench 2.0 | 78.0412 / 49Max (task-specific tools) | 60.7438 / 49Max (task-specific tools) | +17.30 |
| SuperCLUE | 71.812 / 13Reported best (effort unspecified) | 71.294 / 13Reported best (effort unspecified) | +0.52 |
写作与创意表达
GLM-5.3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| Creative Writing | 1,54053 / 118Normal (No Tools) | 2,0776 / 118Normal (No Tools) | -536.80 |
办公与文档流程
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| AutomationBench | 54.806 / 148Max (With Tools) | 48.2017 / 148Max (With Tools) | +6.60 |
文档与图表理解
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| GDP.pdf | 12.8074 / 126Max (No Tools) | 11.2081 / 126Max (No Tools) | +1.60 |
漏洞利用
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| ExploitGym (budget unspecified) | 15.306 / 10Max (With Tools) | 157 / 10Max (With Tools) | +0.30 |
漏洞发现与分析
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| CyberGym | 88.103 / 14Max (With Tools) | 84.505 / 14Max (With Tools) | +3.60 |
科学知识与推理
GLM-5.3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| CritPt | 14.3063 / 207Max (No Tools) | 19.1044 / 207Max (No Tools) | -4.80 |
程序生成与编辑
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| Program Bench | 20.3012 / 18Max (With Tools) | 1913 / 18Max (With Tools) | +1.30 |
综合知识考试
DeepSeek-V4.1-Flash 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| HLE | 63.906 / 247Max (With Tools) | 62.508 / 247Max (With Tools) | +1.40 |
跨应用与工具编排
GLM-5.3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| APEX-Agents | 39.5049 / 63Reported best (effort unspecified) | 56.6021 / 63Reported best (effort unspecified) | -17.10 |
金融
GLM-5.3 领先 1/1| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 | 分差 |
|---|---|---|---|
| Tax Agent Bench | 26.6345 / 72Max (With Tools) | 73.093 / 72Max (With Tools) | -46.46 |
测试模式不同的成绩
共 17 项。两款模型的公开成绩来自不同测试模式，分数不直接可比，不计入胜负和平均分差。
| 评测项 | DeepSeek-V4.1-Flash | GLM-5.3 |
|---|---|---|
| IOI (Vals v2) | 40.28Thinking High (With Tools) | 68.44Max (With Tools) |
| Vals Index | 57.86Thinking High (With Tools) | 56.97Max (With Tools) |
| Vals Index v2.1 | 51.32Thinking High (With Tools) | 53.51Max (task-specific tools) |
| MedCode | 41.17Thinking High (With Tools) | 42.86Max (With Tools) |
| MedScribe | 85.50Thinking High (With Tools) | 88.81Max (With Tools) |
| FrontierCode 1.1 Main | 43.90Thinking High (With Tools) | 40.10Max (With Tools) |
| Vibe Code Bench v1.1 | 84.74Thinking High (With Tools) | 78.12Max (With Tools) |
| SimpleBench | 66.70Historical report (mode unspecified) | 66.20Max (No Tools) |
| LiveBench | 81.11Max (No Tools) | 76.14Reported best (effort unspecified) |
| EMB (Excel Modeling) | 57.21Thinking High (With Tools) | 56.34Max (With Tools) |
| Finance Agent v2 | 53.48Thinking High (With Tools) | 55.84Max (With Tools) |
| Harvey's Legal Agent Benchmark | 6.67Thinking High (With Tools) | 8.33Max (With Tools) |
| Legal Research Bench | 41.35Thinking High (With Tools) | 49.04Max (With Tools) |
| Code Migration | 45.62Thinking High (With Tools) | 44.22Max (With Tools) |
| ProofBench v1.1 (Lean 4) | 54Thinking High (With Tools) | 49Max (With Tools) |
| MysteryMechanism | 21.17Thinking High (With Tools) | 22.97Max (With Tools) |
| Public Benefits Bench v1.1 | 64.28Thinking High (With Tools) | 68.54Max (With Tools) |
规格对比
| 字段 | DeepSeek-V4.1-Flash | GLM-5.3 |
|---|---|---|
| 发布机构 | DeepSeek-AI | 智谱AI |
| 发布时间 | 2026-09-10 | 2026-08-14 |
| 模型类型 | 多模态大模型 | 推理大模型 |
| 架构 | MoE 架构 | MoE 架构 |
| 参数规模 | 5520亿 | 7440亿 |
| 上下文长度 | 1M | 1M |
| 最大输出 | 384K | 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 | DeepSeek-V4.1-Flash | GLM-5.3 |
|---|---|---|
| 文本输入 | ¥1 / 1M tokens | ¥8 / 1M tokens |
| 文本输出 | ¥4 / 1M tokens | ¥28 / 1M tokens |
| 缓存读取 | ¥0.02 / 1M tokens | ¥2 / 1M tokens |
小结
- DeepSeek-V4.1-Flash在以下类目领先:仓库修复与多文件工程 (3/3)、跨能力综合测试 (2/2)、办公与文档流程 (1/1)、文档与图表理解 (1/1)、漏洞利用 (1/1)、漏洞发现与分析 (1/1)、程序生成与编辑 (1/1)、综合知识考试 (1/1)
- GLM-5.3在以下类目领先:数学 (2/3)、自主开发与终端任务 (2/3)、写作与创意表达 (1/1)、科学知识与推理 (1/1)、跨应用与工具编排 (1/1)、金融 (1/1)
- 胶着类目:科学计算与科研编程
23 个同模式 benchmark 上，GLM-5.3 平均高出 24.51 分。
单项差距最大的 benchmark：Creative Writing — DeepSeek-V4.1-Flash 1,540，GLM-5.3 2,077（分差 -536.80）。
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。