Mistral Large 4 vs Mistral Large 4.0 OSS
完整资料:Mistral Large 4(Mistral AI)和 Mistral Large 4.0 OSS(Mistral AI),包括全部基准测试、价格和排名。
更新于 · 数据来自 7 个来源
截至 2026年11月16日,两者共同的基准测试太少,无法判断质量高下。看重知识(41 比 34)选 Large 4;看重推理(64 比 46)和数学(66 比 54)选 Large 4.0 OSS。
要点(截至 2026年11月16日)
- Mistral Large 4 每百万 token 混合价格为 $1.05(标价,2026年11月16日)。
按用途怎么选
推荐模型的数值列在前面。类别得分采用 BenchLeader 标尺(50 为模型平均水平,15 分为一个标准差)。领先半个标准差及以上为“高”置信度;相差不到 1 分则视为“均可”。
主要差异
- 共同的基准测试太少,无法给出结论。
- Mistral Large 4 在知识方面更强。
- Mistral Large 4.0 OSS (high) 在数学、推理方面更强。
概览
| 规格 | Mistral Large 4 | Mistral Large 4.0 OSS |
|---|---|---|
| 提供商 | Mistral AI | Mistral AI |
| 发布日期 | 2026年10月6日 | 2026年10月29日 |
| 上下文窗口(token) | 1,048,576 | – |
| 对比时的配置 | Mistral Large 4 | Mistral Large 4.0 OSS (high) |
| 开放权重 | 是 | 否 |
| 每百万 token 价格(输入 / 输出 / 缓存输入) | $0.700 / $2.10 / $0.070 | – |
逐类别对比
类别得分采用 BenchLeader 标尺:50 为参评模型的平均水平,15 分为一个标准差。差距不到 1 分视为持平。
| 类别 | Mistral Large 4 | Mistral Large 4.0 OSS (high) | 领先者 |
|---|---|---|---|
| 推理 | 46.2 | 64.2 | Mistral Large 4.0 OSS |
| 编程 | 53.8 | – | 无法比较 |
| 智能体与工具 | 48.8 | – | 无法比较 |
| 数学 | 54.2 | 66.4 | Mistral Large 4.0 OSS |
| 知识 | 40.7 | 34.1 | Mistral Large 4 |
| 人类偏好 | 59.7 | – | 无法比较 |
| 多模态 | 58.0 | – | 无法比较 |
| 长上下文 | 65.3 | – | 无法比较 |
同一图表对比
各领域采用与 Index 相同的标尺,平均模型为 50 分。单击某个领域可查看构成它的基准测试。 可从列表中最多选择 5 个模型。
未绘制(测得的模型少于两个):编程, 智能体与工具, 指令遵循, 长上下文, 多模态, 人类偏好。
价格、速度与延迟
标价按每 1 个输出 token 对应 3 个输入 token 混合计算。每项任务的价格采用方法论页面中的 token 数量。每项基准测试任务的成本、token 数和耗时均为实测值,取自 Artificial Analysis 整个 Intelligence Index 测试的平均值。
| 指标 | Mistral Large 4 | Mistral Large 4.0 OSS (high) |
|---|---|---|
| 每百万 token 混合价格 | $1.05 | – |
| 每百万 token 输入 / 输出 | $0.700 / $2.10 | – / – |
| 每百万 token 缓存输入 | $0.070 | – |
| 每项任务价格:聊天回复 | $0.0009 | – |
| 每项任务价格:总结一份 30 页的报告 | $0.0097 | – |
| 每项任务价格:代码编辑 | $0.0074 | – |
| 每项任务价格:智能体编程会话 | $0.050 | – |
| 每项任务价格:结构化信息提取 | $0.0018 | – |
| 智能体编程会话(输入大多命中缓存) | $0.022 | – |
| 每项基准测试任务成本(AA Index) | $4.21 | – |
| 每项任务输出 token(AA Index) | 135k | – |
| 其中推理部分 | 102k | – |
| 每项任务耗时(AA Index) | 23 min | – |
| 输出速度 | 97 tok/s | – |
| 首次回答时间 | 40.9 秒 | – |
| 完整回答耗时(最近一次测量 2026年11月13日) | 119.0 秒 | – |
每个模型都按其排名最高的推理强度设置进行对比。Mistral Large 4 得分 52.9,首个回答 token 需 41 秒;Mistral Large 4 (high) 为 48.3(45 秒),Mistral Large 4 (medium) 为 45.9(27 秒)。Mistral Large 4.0 OSS (high)。推理强度越高,回答前思考越久:得分更高,但等待更长、计费 token 也更多。 推理强度如何影响得分和成本
全部基准测试
共 111 项基准测试,按类别分组;点击名称可查看该排行榜及领先模型。† 表示同一模型在另一推理强度设置下的测量值,仅供参考,不计入胜场。
| 指标 | 发布方 | Mistral Large 4 | Mistral Large 4.0 OSS (high) |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 52.9 | – |
| 智能体与工具得分 | BenchLeader | 48.8 | – |
| 编程得分 | BenchLeader | 53.8 | – |
| 人类偏好得分 | BenchLeader | 59.7 | – |
| 知识得分 | BenchLeader | 40.7 | 34.1 |
| 长上下文得分 | BenchLeader | 65.3 | – |
| 数学得分 | BenchLeader | 54.2 | 66.4 |
| 多模态得分 | BenchLeader | 58.0 | – |
| 推理得分 | BenchLeader | 46.2 | 64.2 |
| 混合价格(美元/百万 token) | BenchLeader | $1.05 | – |
| 输出速度 | BenchLeader | 97 tok/s | – |
| 首次回答时间 | BenchLeader | 40.9 秒 | – |
| 上下文窗口 | BenchLeader | 1,048,576 token | – |
| 推理 | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 88.1%† | 90.2% |
| Arena Hard Prompts | Arena (LMArena) | 1453 | – |
| LiveBench 2.0 Reasoning | LiveBench | 85.2%† | – |
| LiveBench 1.0 Reasoning | LiveBench | 83.9%† | – |
| Humanity's Last Exam (AA) | Artificial Analysis | 35.0% | – |
| HLE-Diamond (AA) | Artificial Analysis | 15.1% | – |
| CritPt | Artificial Analysis | 10.6% | – |
| CritPt v2 hard subset (AA) | Artificial Analysis | 24.5% | – |
| MysteryMechanism | Vals AI | 12.2%† | – |
| Arena Expert | Arena (LMArena) | 1464 | – |
| Chess Puzzles | Epoch AI Benchmarking Hub | 33.0%† | 29.0% |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 17.0%† | 20.0% |
| LMCA | Epoch AI Benchmarking Hub | 35.8% | – |
| DTBench | Epoch AI Benchmarking Hub | 88.2% | – |
| 编程 | |||
| SciCode | SciCode | 54.2% | – |
| FrontierCode | Cognition | 39.4%† | – |
| Arena Coding | Arena (LMArena) | 1484 | – |
| Arena WebDev | Arena (LMArena) | 1565 | – |
| LiveBench 2.0 Coding | LiveBench | 79.4%† | – |
| LiveBench 1.0 Coding | LiveBench | 77.2%† | – |
| IOI | Vals AI | 45.3%† | – |
| Code Migration | Vals AI | 30.6%† | – |
| SWE-bench Multimodal | Vals AI | 33.1%† | – |
| Vibe Code Bench 1-100 | Vals AI | 14.3%† | – |
| Vibe Code Bench v1.1 | Vals AI | 78.4%† | – |
显示全部 111 项基准测试
| 指标 | 发布方 | Mistral Large 4 | Mistral Large 4.0 OSS (high) |
|---|---|---|---|
| 编程 | |||
| SWE Atlas v1.1 | Scale AI SEAL | 45.2% | – |
| SWE Atlas v1.1: Codebase QnA | Scale AI SEAL | 61.3% | – |
| SWE Atlas v1.1: Test Writing | Scale AI SEAL | 40.0% | – |
| SWE Atlas v1.1: Refactoring | Scale AI SEAL | 34.3% | – |
| MirrorCode | Epoch AI Benchmarking Hub | 40.0%† | – |
| Surface Evolver Bench | Epoch AI Benchmarking Hub | 28.1% | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 1213.2 | – |
| WeirdML v3 | WeirdML | 4.5%† | – |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 26.8% | – |
| GameBench 2 by reasoning setting | DuelLab GameBench 2 | 57.6† | – |
| Design Arena Website | Design Arena | 1258 | – |
| FrontierCode 1.1 (AA) | Artificial Analysis | 41.5% | – |
| FrontierCode 2.0 (AA) | Artificial Analysis | 13.6% | – |
| 智能体与工具 | |||
| Terminal-Bench | Terminal-Bench | 27.6%† | – |
| APEX-Agents | Mercor | 51.6%† | – |
| Arena Agent | Arena (LMArena) | -5.4 | – |
| LiveBench Long Agentic | LiveBench | 24.7%† | – |
| LiveBench Browser | LiveBench | 17.4%† | – |
| LiveBench Agentic Coding | LiveBench | 57.2%† | – |
| GDPval-AA v2.1, normalised score | Artificial Analysis | 46.1% | – |
| GDPval-AA v2.1 | Artificial Analysis | 1431 | – |
| Finance Agent v2 | Vals AI | 54.7%† | – |
| Finance Agent v2.1 | Vals AI | 61.4%† | – |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | 17.5%† | – |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 15.8%† | – |
| Legal Research Bench | Vals AI | 31.7%† | – |
| Tax Agent Bench | Vals AI | 30.4%† | – |
| Terminal-Bench 4.0 (Vals) | Vals AI | 27.8%† | – |
| Arena Agent Task Completion | Arena (LMArena) | -4.7 | – |
| AutomationBench | Artificial Analysis | 59.9% | – |
| GDP.pdf (AA) | Artificial Analysis | 18.6% | – |
| Harvey LAB-AA v1.1 | Harvey | 3.1% | – |
| AA-Briefcase v1.1 | Artificial Analysis | 1393 | – |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | 4.8% | – |
| CWE-Bench (AA) | Artificial Analysis | 50.8% | – |
| DeepSec-Bench (AA) | Artificial Analysis | 16.0% | – |
| CyberGym end-to-end (AA) | Artificial Analysis | 81.7% | – |
| MKA (AA) | Artificial Analysis | 35.1%† | – |
| 数学 | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 58.6%† | – |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 26.8%† | – |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 98.3%† | 99.4% |
| ProofBench | Vals AI | 10.0% | – |
| LiveBench Mathematics | LiveBench | 93.6%† | – |
| Arena Maths | Arena (LMArena) | 1459 | – |
| 知识 | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 20.0%† | 22.4% |
| LiveBench 2.0 Data Analysis | LiveBench | 76.6%† | – |
| LiveBench Research | LiveBench | 80.3%† | – |
| LiveBench 1.0 Data Analysis | LiveBench | 76.5%† | – |
| AA-Omniscience v1.1 | Artificial Analysis | -7.5 | – |
| BioMysteryBench | Vals AI | 67.0%† | – |
| Excel Modeling Benchmark | Vals AI | 56.2%† | – |
| Excel Modeling Benchmark v1.1 | Vals AI | 69.0%† | – |
| MedCode | Vals AI | 40.7%† | – |
| MedScribe | Vals AI | 80.4%† | – |
| Immigration Bench | Vals AI | 63.7%† | – |
| EBR-bench | Epoch AI Benchmarking Hub | – | 18.1% |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | 23.9% | – |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | 58.7% | – |
| 指令遵循 | |||
| LiveBench Language | LiveBench | 49.6%† | – |
| LiveBench Instruction Following | LiveBench | 64.8%† | – |
| Arena Instruction Following | Arena (LMArena) | 1418 | – |
| 人类偏好 | |||
| Arena Text | Arena (LMArena) | 1420 | – |
| Arena Creative Writing | Arena (LMArena) | 1359 | – |
| Arena Multi-turn | Arena (LMArena) | 1429 | – |
| Arena Longer Queries | Arena (LMArena) | 1420 | – |
| 多模态 | |||
| Arena Vision | Arena (LMArena) | 1244 | – |
| LiveBench Design | LiveBench | 76.3%† | – |
| MMMU-Pro | Artificial Analysis | 76.4% | – |
| Furniture Assembly | Epoch AI Benchmarking Hub | 20.8%† | 30.8% |
| VIALS-AA | Handshake AI | 20.5%† | – |
| Code2Video-AA | Artificial Analysis | 931 | – |
| AA-3D | Artificial Analysis | 882 | – |
| 长上下文 | |||
| AA-LCR v1.1 | Artificial Analysis | 81.3% | – |
| MLCR | Artificial Analysis | 7.8% | – |
| 综合 | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 154.6 | – |
| LiveBench 2.0 | LiveBench | 60.3%† | – |
| LiveBench 1.0 | LiveBench | 71.8%† | – |
| AA Intelligence Index v5.0.1 | Artificial Analysis | 25.3 | – |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 38.4 | – |
| AA Progress Index v1 | Artificial Analysis | 312.4 | – |
| Vals Finance Index | Vals AI | 56.9%† | – |
| Vals Legal Index | Vals AI | 65.6%† | – |
| AA Cyber Index | Artificial Analysis | 49.5 | – |
| 安全与诚实 | |||
| FORTRESS | Scale AI SEAL | 23.1% | – |
| DistressBench | Scale AI SEAL | 77.0% | – |
| 元数据 | |||
| AA Openness Index | Artificial Analysis | 38.9† | – |
数据截至 2026年11月16日。对比的是每个模型的最佳配置;每项分数的来源链接都在模型页面上。
DuelLab GameBench 2, release gb2-20261111-7b598bb13de4 (https://benchmarks.duellab.org/releases/gb2-20261111-7b598bb13de4/), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Scores mapped to BenchLeader model names; not endorsed by DuelLab.
变化
- Mistral Large 4:BenchLeader Index 52.2 → 52.9。
- Mistral Large 4:1 项新结果(GDPval-AA v2.1)。
- Mistral Large 4:1 项分数修订。
- Mistral Large 4:2 项分数修订。
- Mistral Large 4:6 项分数修订。
- Mistral Large 4:BenchLeader Index 52.7 → 52.2。
- Mistral Large 4:7 项分数修订。
- Mistral Large 4:8 项分数修订。
- Mistral Large 4:2 项分数修订。
- Mistral Large 4:1 项分数修订。
相关对比
- Mistral Large 4 vs Muse Spark 1.3
- Mistral Large 4 vs MiniMax M3.1 Flash
- Mistral Large 4 vs Muse Ember
- Mistral Large 4.0 OSS vs Inkling
- Mistral Large 4.0 OSS vs Nemotron 3.5 Super VL 120B A12B
- Mistral Large 4.0 OSS vs Qwen3.5-35B-A3B
排行榜: 综合最佳大模型 · 最适合智能体的大模型 · 便宜好用的大模型
Mistral Large 4 vs Mistral Large 4.0 OSS:常见问题
- Mistral Large 4.0 OSS 是什么时候发布的?
- Mistral Large 4.0 OSS 于 2026年10月29日 发布。