跳转至

Mainstream Models Overview | 主流大模型全景

The global AI model landscape is evolving rapidly. This document summarizes key information about mainstream models to help you quickly grasp their characteristics and application scenarios.

当前全球 AI 大模型发展迅速。本文档汇总了主流模型的关键信息,帮助您快速把握其核心特点和应用场景。

Global Mainstream Models Summary | 全球主流模型汇总

Model Name (模型名称) Organization (所属机构) Core Features (核心特点) Main Applications (主要应用领域) Hardware Requirements (关键硬件资源需求)
o1 Series (o1-preview/o1-mini) OpenAI (USA) Specialized in complex reasoning ("System 2" thinking). Uses Chain of Thought reinforcement learning. Strong in Math/Code.
专精于复杂推理(“系统 2”思维)。采用思维链强化学习训练。数学和编程能力极强。
STEM research, complex coding, advanced math problem solving.
STEM 科研、复杂代码开发、高阶数学解题。
API only.
目前仅通过 API 调用。
GPT-4o OpenAI (USA) Native multimodal (text, image, audio), extremely fast. "Omni" model.
原生多模态(文本、图像、音频),响应极快。“全能”模型。
Real-time voice interaction, vision analysis, general assistant.
实时语音交互、视觉分析、通用助手。
API only.
通过 API 调用。
Claude 3.5 Sonnet Anthropic (USA) Best-in-class coding and reasoning. Balances speed and intelligence perfectly. "Artifacts" UI integration.
编程与推理能力目前业界顶尖。完美平衡速度与智能。“Artifacts”界面集成体验佳。
Software development, data analysis, creative writing.
软件开发、数据分析、创意写作。
API only.
通过 API 调用。
Gemini 1.5 Pro Google DeepMind (USA) Massive context window (up to 2M tokens). Deep ecosystem integration (Workspace).
超大上下文窗口(可达 200 万 token)。深度集成 Google 生态(Workspace)。
Long-context understanding (books, codebases), multimodal analysis.
长文本理解(整书、代码库)、多模态分析。
API or Google Cloud.
API 或 Google Cloud。
Grok 2 xAI (USA) Integrated with X (Twitter) real-time data. Unfiltered/Fun mode. Image generation via FLUX.
集成 X (Twitter) 实时数据。具备“无过滤/有趣”模式。通过 FLUX 生成图像。
Real-time news analysis, social media interaction, image generation.
实时新闻分析、社交媒体交互、图像生成。
API or X Premium.
API 或 X Premium 订阅。
Llama 3.2 Meta (USA) Open weights. 1B/3B (Edge) and 11B/90B (Vision-Multimodal). Standard for open-source.
开放权重。1B/3B(边缘侧)和 11B/90B(视觉多模态)。开源领域的标准。
Edge devices (mobile), custom enterprise apps, visual reasoning.
边缘设备(手机)、企业定制应用、视觉推理。
Consumer GPUs (RTX 4090) to H100s depending on size.
视尺寸而定,从消费级显卡到 H100。
Mistral Large 2 Mistral AI (France) Strong reasoning, coding, and multilingual support. Direct competitor to GPT-4o/Llama 3.1 405B.
强大的推理、编程及多语言支持。GPT-4o 和 Llama 3.1 405B 的有力竞争者。
Enterprise RAG, multilingual tasks, coding agents.
企业级 RAG、多语言任务、编程智能体。
High-end GPU clusters.
高端 GPU 集群。
DeepSeek-V3 (深度求索) DeepSeek (China) MoE Architecture (671B params, 37B active). Strong coding/math. Top-tier open weights.
MoE 架构(671B 参数,37B 激活)。编程/数学能力极强。开源界顶流。
Coding, Math, Scientific research, cost-effective inference.
编程、数学计算、科研、高性价比推理。
High-end Consumer (Quantized) to H800 clusters.
高端消费级(量化版)到 H800 集群。
Qwen 2.5 (通义千问) Alibaba (China) Comprehensive size range (0.5B to 72B+). "Coder" and "Math" specialized variants.
尺寸覆盖全(0.5B 到 72B+)。拥有“Coder”和“Math”专项版。
General purpose, coding assistants, mobile deployment.
通用场景、编程助手、移动端部署。
Consumer to Pro cards.
消费级到专业卡。
Yi-Lightning 01.AI (China) Fast inference, strong performance on LMSYS leaderboard.
推理速度快,LMSYS 榜单表现强劲。
High-throughput chat applications.
高吞吐量对话应用。
API mostly.
主要为 API。
Hunyuan Video (混元视频) Tencent (China) Open source video generation foundation model. High quality.
开源视频生成基础模型。画质高。
Video content creation, advertising.
视频内容创作、广告。
High VRAM GPUs.
高显存 GPU。

Note: Model versions and capabilities evolve rapidly. This table represents a snapshot of the landscape. 注:模型版本和能力迭代迅速,本表格仅代表当前阶段的行业概览。