Agent Economy

Google 发布 Gemini 3.1 Pro：新一代多模态推理模型

Google Releases Gemini 3.1 Pro: Next-Gen Multimodal Reasoning Model

Google DeepMind 发布 Gemini 3.1 Pro 模型卡，这是 Gemini 3 系列的最新迭代，也是 Google 目前最先进的复杂任务处理模型。作为原生多模态推理模型，Gemini 3.1 Pro 能够处理来自文本、音频、图像、视频和完整代码库的大规模多模态信息。

核心规格

  • 上下文窗口：100 万 tokens
  • 输出长度：64K tokens
  • 输入类型：文本、图像、音频、视频
  • 基础模型：基于 Gemini 3 Pro 架构

性能表现

Gemini 3.1 Pro 在多个关键基准测试中显著超越 Gemini 3 Pro 和竞争对手：

学术推理

  • Humanity's Last Exam：44.4%（无工具）/ 51.4%（搜索+代码）- 业界领先
  • ARC-AGI-2：77.1% - 抽象推理难题
  • GPQA Diamond：94.3% - 科学知识问答

编程能力

  • SWE-Bench Verified：80.6% - 智能体编程
  • Terminal-Bench 2.0：68.5% - 终端代码代理
  • LiveCodeBench Pro：Elo 2887 - 竞争性编程

多模态与长上下文

  • MMMU-Pro：80.5% - 多模态理解
  • MRCR v2 (128k)：84.9% - 长上下文检索
  • MMMLU：92.6% - 多语言问答

智能体能力

  • APEX-Agents：33.5% - 长周期专业任务
  • BrowseComp：85.9% - 智能体搜索
  • MCP Atlas：69.2% - 多步工作流

应用场景

Gemini 3.1 Pro 特别适合以下应用：

  • 智能体性能任务
  • 高级编程和代码生成
  • 长上下文和多模态理解
  • 算法开发
  • 复杂问题求解和战略规划

阅读完整模型卡：https://deepmind.google/models/model-cards/gemini-3-1-pro/

Google DeepMind has released the Gemini 3.1 Pro model card, the latest iteration in the Gemini 3 series and Google's most advanced model for complex tasks. As a natively multimodal reasoning model, Gemini 3.1 Pro can process massively multimodal information from text, audio, images, video, and entire code repositories.

Core Specifications

  • Context Window: 1 million tokens
  • Output Length: 64K tokens
  • Input Types: Text, images, audio, video
  • Base Architecture: Based on Gemini 3 Pro

Performance

Gemini 3.1 Pro significantly outperforms Gemini 3 Pro and competitors across key benchmarks:

Academic Reasoning

  • Humanity's Last Exam: 44.4% (no tools) / 51.4% (search+code) - Industry leading
  • ARC-AGI-2: 77.1% - Abstract reasoning puzzles
  • GPQA Diamond: 94.3% - Scientific knowledge Q&A

Coding Capabilities

  • SWE-Bench Verified: 80.6% - Agentic coding
  • Terminal-Bench 2.0: 68.5% - Terminal code agent
  • LiveCodeBench Pro: Elo 2887 - Competitive programming

Multimodal & Long Context

  • MMMU-Pro: 80.5% - Multimodal understanding
  • MRCR v2 (128k): 84.9% - Long context retrieval
  • MMMLU: 92.6% - Multilingual Q&A

Agentic Capabilities

  • APEX-Agents: 33.5% - Long horizon professional tasks
  • BrowseComp: 85.9% - Agentic search
  • MCP Atlas: 69.2% - Multi-step workflows

Use Cases

Gemini 3.1 Pro is particularly well-suited for:

  • Agentic performance tasks
  • Advanced coding and code generation
  • Long context and multimodal understanding
  • Algorithm development
  • Complex problem solving and strategic planning

Read the full model card at https://deepmind.google/models/model-cards/gemini-3-1-pro/

GoogleGeminiAI-ModelMultimodalReasoningAgentic-AI
← All articles