趋近智
GPT-OSS 120B
活跃参数
117B
上下文长度
128K
模态
Text
架构
Mixture of Experts (MoE)
许可证
Apache 2.0
发布日期
5 Aug 2025
训练数据截止日期
Jun 2024
技术规格
专家参数总数
5.1B
专家数量
128
活跃专家
4
注意力结构
Multi-Head Attention
隐藏维度大小
2880
层数
36
注意力头
-
键值头
-
激活函数
SwigLU
归一化
RMS Normalization
位置嵌入
Absolute Position Embedding
系统要求
不同量化方法和上下文大小的显存要求
GPT-OSS 120B
GPT-OSS 120B is a large open-weight model from OpenAI, designed to operate in data centers and on high-end desktops and laptops. It is developed to support advanced reasoning, agentic tasks, and diverse developer use cases, functioning as a text-only model for both input and output modalities.
评估基准
排名
#86
| 基准 | 分数 | 排名 |
|---|---|---|
Summarization ProLLM Summarization | 0.98 | 🥉 3 |
General Knowledge MMLU | 0.90 | 🥉 3 |
StackUnseen ProLLM Stack Unseen | 0.93 | 4 |
Professional Knowledge MMLU Pro | 0.81 | 20 |
Coding Aider Coding | 0.42 | 28 |
Mathematics LiveBench Mathematics | 0.69 | 30 |
Agentic Coding LiveBench Agentic | 0.17 | 40 |
Graduate-Level QA GPQA | 0.81 | 44 |
Web Development WebDev Arena | 1092.96 | 45 |
Coding LiveBench Coding | 0.60 | 49 |
Reasoning LiveBench Reasoning | 0.39 | 50 |
Data Analysis LiveBench Data Analysis | 0.57 | 50 |