e:[["$","link",null,{"rel":"alternate","hrefLang":"zh-CN","href":"https://www.datalearner.com/ai-models/pretrained-models/openai-o1-mini/analysis"}],["$","link",null,{"rel":"alternate","hrefLang":"zh","href":"https://www.datalearner.com/ai-models/pretrained-models/openai-o1-mini/analysis"}],["$","link",null,{"rel":"alternate","hrefLang":"x-default","href":"https://www.datalearner.com/ai-models/pretrained-models/openai-o1-mini/analysis"}]] a:[["$","meta","0",{"charSet":"utf-8"}],["$","meta","1",{"name":"viewport","content":"width=device-width, initial-scale=1, maximum-scale=5"}],["$","meta","2",{"name":"theme-color","media":"(prefers-color-scheme: light)","content":"#f8fafc"}],["$","meta","3",{"name":"theme-color","media":"(prefers-color-scheme: dark)","content":"#0f172a"}]] 15:T4e4,{"@context":"https://schema.org","@graph":[{"@type":"CollectionPage","@id":"https://www.datalearner.com/en/ai-models/pretrained-models/openai-o1-mini/analysis#webpage","url":"https://www.datalearner.com/en/ai-models/pretrained-models/openai-o1-mini/analysis","headline":"OpenAI o1-mini Benchmark Results & Model Comparisons","description":"OpenAI o1-mini currently shows benchmark results led by HumanEval (4 / 39, score 92.40), MMLU Pro (49 / 116, score 80.30), MMLU (39 / 64, score 85.20).","author":{"@type":"Organization","name":"DataLearnerAI","url":"https://www.datalearner.com"},"publisher":{"@type":"Organization","name":"DataLearnerAI","logo":{"@type":"ImageObject","url":"https://www.datalearner.com/dl_logo_rect.webp"}},"mainEntity":{"@id":"https://www.datalearner.com/en/ai-models/pretrained-models/openai-o1-mini#software"},"mainEntityOfPage":{"@type":"WebPage","@id":"https://www.datalearner.com/en/ai-models/pretrained-models/openai-o1-mini/analysis#webpage"},"hasPart":[]},{"@type":"SoftwareApplication","@id":"https://www.datalearner.com/en/ai-models/pretrained-models/openai-o1-mini#software","name":"OpenAI o1-mini","url":"https://www.datalearner.com/en/ai-models/pretrained-models/openai-o1-mini","applicationCategory":"AI Model"}]}17:T9a4,

OpenAI于2024年9月12日推出了o1-mini模型，这是一款专为STEM领域（科学、技术、工程和数学）设计的高性价比推理模型。作为o1-preview模型的经济高效替代方案，o1-mini在多个评估基准上表现出色。

主要特点：

STEM推理优化： o1-mini专注于需要强大推理能力的任务，特别是在数学和编程方面。在美国数学邀请赛（AIME）和Codeforces等基准测试中，o1-mini的表现几乎与o1-preview相当。
成本与速度优势： o1-mini的成本比o1-preview低约80%，为需要强大推理能力但不需要广泛世界知识的应用提供了经济实惠的解决方案。此外，o1-mini具有更高的速率限制和更低的延迟，适合实时应用。

基准测试表现：

数学能力：在AIME测试中，o1-mini取得了70.0%的得分，接近o1的74.4%，远超o1-preview的44.6%。这一成绩相当于美国前500名高中生的水平。
编程能力：在Codeforces平台上，o1-mini的Elo评分达到1650，接近o1的1673，显著高于o1-preview的1258。这意味着o1-mini位于竞争性程序员的前86%。
STEM推理：在MATH-500基准测试中，o1-mini在零样本链式思维提示下达到了90.0%的准确率，展现了强大的数学问题解决能力。

技术规格：

上下文窗口： o1-mini支持128,000个token的输入上下文窗口，能够在单次请求中处理大量信息。
输出容量：模型每次请求最多可生成65,500个token，确保提供详细且全面的响应。

安全性与一致性：

o1-mini采用与o1-preview相同的对齐和安全技术。内部评估显示，与GPT-4o相比，o1-mini对越狱尝试的抵抗力提高了59%，确保遵守安全指南，降低生成有害内容的风险。

局限性：

尽管o1-mini在STEM相关

OpenAI o1-mini Benchmark Details

Benchmark Results

Benchmark Results

数学推理

OpenAI o1-mini Benchmark Details

Benchmark Results

Benchmark Results

数学推理