GLM-5：开源Agentic Engineering的新纪元

✨步子哥 (steper) • 2026年02月21日 11:55

                        <!DOCTYPE html><html lang="zh-CN"><head>
    <meta charset="UTF-8"/>
    <meta name="viewport" content="width=device-width, initial-scale=1.0"/>
    <title>GLM-5 技术深度分析报告</title>
    <script src="https://cdn.tailwindcss.com"></script>
    <script src="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/js/all.min.js"></script>
    <link href="https://fonts.googleapis.com/css2?family=Crimson+Text:ital,wght@0,400;0,600;1,400&amp;family=Inter:wght@300;400;500;600;700&amp;display=swap" rel="stylesheet"/>
    <script src="https://cdn.jsdelivr.net/npm/mermaid@10.6.1/dist/mermaid.min.js"></script>
    <style>
        :root {
            --primary: #1e293b;
            --secondary: #334155;
            --accent: #0f172a;
            --muted: #64748b;
            --background: #f8fafc;
            --surface: #ffffff;
            --border: #e2e8f0;
        }
        
        body {
            font-family: 'Inter', sans-serif;
            background: var(--background);
            color: var(--primary);
            line-height: 1.7;
            overflow-x: hidden;
        }
        
        .serif {
            font-family: 'Crimson Text', serif;
        }
        
        .toc-fixed {
            position: fixed;
            left: 0;
            top: 0;
            width: 180px;
            height: 100vh;
            background: var(--surface);
            border-right: 1px solid var(--border);
            z-index: 1000;
            overflow-y: auto;
            padding: 2rem 1.5rem;
        }
        
        .main-content {
            margin-left: 180px;
            min-height: 100vh;
        }
        
        .hero-grid {
            display: grid;
            grid-template-columns: 1fr 1fr;
            gap: 2rem;
            align-items: center;
            min-height: 60vh;
        }
        
        .hero-text {
            background: linear-gradient(135deg, var(--primary) 0%, var(--secondary) 100%);
            -webkit-background-clip: text;
            -webkit-text-fill-color: transparent;
            background-clip: text;
        }
        
        .citation {
            color: var(--muted);
            text-decoration: none;
            font-size: 0.875rem;
            vertical-align: super;
            margin-left: 2px;
        }
        
        .citation:hover {
            color: var(--accent);
        }
        
        .highlight-box {
            background: linear-gradient(135deg, #f1f5f9 0%, #e2e8f0 100%);
            border-left: 4px solid var(--accent);
            padding: 1.5rem;
            margin: 2rem 0;
            border-radius: 0 8px 8px 0;
        }
        
        .architecture-diagram {
            background: var(--surface);
            border: 2px solid var(--border);
            border-radius: 12px;
            padding: 2rem;
            margin: 2rem 0;
            box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1);
        }
        
        .performance-card {
            background: var(--surface);
            border: 1px solid var(--border);
            border-radius: 8px;
            padding: 1.5rem;
            margin: 1rem 0;
            transition: all 0.3s ease;
        }
        
        .performance-card:hover {
            box-shadow: 0 10px 25px -5px rgb(0 0 0 / 0.1);
            transform: translateY(-2px);
        }
        
        .toc-item {
            display: block;
            padding: 0.5rem 0;
            color: var(--muted);
            text-decoration: none;
            font-size: 0.875rem;
            transition: color 0.2s ease;
        }
        
        .toc-item:hover, .toc-item.active {
            color: var(--accent);
            font-weight: 500;
        }
        
        .toc-item.sub {
            padding-left: 1rem;
            font-size: 0.8rem;
        }
        
        .toc-toggle {
            display: none;
            position: fixed;
            top: 1rem;
            left: 1rem;
            z-index: 1001;
            background: var(--surface);
            border: 1px solid var(--border);
            border-radius: 4px;
            padding: 0.5rem;
            cursor: pointer;
        }
        
        <span class="mention-invalid">@media</span> (max-width: 1024px) {
            .toc-fixed {
                transform: translateX(-100%);
                transition: transform 0.3s ease;
            }
            
            .toc-fixed.active {
                transform: translateX(0);
            }
            
            .main-content {
                margin-left: 0;
            }
            
            .hero-grid {
                grid-template-columns: 1fr;
                gap: 1rem;
            }
            
            .toc-toggle {
                display: block;
            }
            
            body.toc-active {
                overflow: hidden;
            }
        }
        
        <span class="mention-invalid">@media</span> (max-width: 768px) {
            .mermaid-control-btn:not(.reset-zoom) {
                display: none;
            }
            .mermaid-controls {
                top: auto;
                bottom: 15px;
                right: 15px;
            }
        }
        
        /* Mermaid chart styling */
        .mermaid-container {
            display: flex;
            justify-content: center;
            min-height: 300px;
            max-height: 800px;
            background: #ffffff;
            border: 2px solid #e5e7eb;
            border-radius: 12px;
            padding: 30px;
            margin: 30px 0;
            box-shadow: 0 8px 25px rgba(0, 0, 0, 0.08);
            position: relative;
            overflow: hidden;
        }

        .mermaid-container .mermaid {
            width: 100%;
            max-width: 100%;
            height: 100%;
            cursor: grab;
            transition: transform 0.3s ease;
            transform-origin: center center;
            display: flex;
            justify-content: center;
            align-items: center;
            touch-action: none;
            -webkit-user-select: none;
            -moz-user-select: none;
            -ms-user-select: none;
            user-select: none;
        }

        .mermaid-container .mermaid svg {
            max-width: 100%;
            height: 100%;
            display: block;
            margin: 0 auto;
        }

        .mermaid-container .mermaid:active {
            cursor: grabbing;
        }

        .mermaid-container.zoomed .mermaid {
            height: 100%;
            width: 100%;
            cursor: grab;
        }

        .mermaid-controls {
            position: absolute;
            top: 15px;
            right: 15px;
            display: flex;
            gap: 10px;
            z-index: 20;
            background: rgba(255, 255, 255, 0.95);
            padding: 8px;
            border-radius: 8px;
            box-shadow: 0 2px 8px rgba(0, 0, 0, 0.1);
        }

        .mermaid-control-btn {
            background: #ffffff;
            border: 1px solid #d1d5db;
            border-radius: 6px;
            padding: 10px;
            cursor: pointer;
            transition: all 0.2s ease;
            color: #374151;
            font-size: 14px;
            min-width: 36px;
            height: 36px;
            text-align: center;
            display: flex;
            align-items: center;
            justify-content: center;
        }

        .mermaid-control-btn:hover {
            background: #f8fafc;
            border-color: #3b82f6;
            color: #3b82f6;
            transform: translateY(-1px);
        }

        .mermaid-control-btn:active {
            transform: scale(0.95);
        }
    </style>
  <base target="_blank">
</head>

  <body>
    <!-- Toggle button for TOC on mobile -->
    <button class="toc-toggle" id="tocToggle">
      <i class="fas fa-bars"></i>
    </button>

    <!-- Fixed Table of Contents -->
    <nav class="toc-fixed" id="tocNav">
      <div class="mb-6">
        <h3 class="serif text-lg font-semibold text-gray-900 mb-4">目录</h3>
      </div>

      <div class="space-y-1">
        <a href="#overview" class="toc-item">模型概述与核心定位</a>
        <a href="#paradigm-shift" class="toc-item sub">从&#34;Vibe Coding&#34;到&#34;Agentic Engineering&#34;</a>
        <a href="#open-source-positioning" class="toc-item sub">开源SOTA定位</a>
        <a href="#architecture" class="toc-item">模型架构创新</a>
        <a href="#moe-architecture" class="toc-item sub">混合专家架构</a>
        <a href="#dsa-mechanism" class="toc-item sub">DeepSeek稀疏注意力</a>
        <a href="#mtp-optimization" class="toc-item sub">多Token预测</a>
        <a href="#training" class="toc-item">训练方法与基础设施</a>
        <a href="#pretraining" class="toc-item sub">预训练策略</a>
        <a href="#slime-framework" class="toc-item sub">Slime异步RL框架</a>
        <a href="#performance" class="toc-item">性能表现与基准测试</a>
        <a href="#academic-benchmarks" class="toc-item sub">学术基准测试</a>
        <a href="#real-world" class="toc-item sub">真实场景能力</a>
        <a href="#deployment" class="toc-item">工程实现与部署优化</a>
        <a href="#token-efficiency" class="toc-item sub">Token效率优化</a>
        <a href="#hardware" class="toc-item sub">硬件适配</a>
        <a href="#evolution" class="toc-item">技术演进与架构溯源</a>
        <a href="#limitations" class="toc-item">研究局限与未来方向</a>
      </div>
    </nav>

    <!-- Main Content -->
    <main class="main-content">
      
      <!-- Key Highlights -->
      <section class="py-12 px-8 bg-white">
        <div class="max-w-6xl mx-auto">
          <div class="grid grid-cols-1 md:grid-cols-3 gap-8">
            <div class="performance-card">
              <div class="flex items-center mb-4">
                <i class="fas fa-microchip text-2xl text-blue-600 mr-3"></i>
                <h3 class="serif text-lg font-semibold">架构创新</h3>
              </div>
              <div class="space-y-3">
                <div class="flex justify-between">
                  <span class="text-gray-600">总参数</span>
                  <span class="font-semibold">744B</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">激活参数</span>
                  <span class="font-semibold">40B/44B</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">专家数量</span>
                  <span class="font-semibold">256个</span>
                </div>
              </div>
            </div>

            <div class="performance-card">
              <div class="flex items-center mb-4">
                <i class="fas fa-bolt text-2xl text-green-600 mr-3"></i>
                <h3 class="serif text-lg font-semibold">效率突破</h3>
              </div>
              <div class="space-y-3">
                <div class="flex justify-between">
                  <span class="text-gray-600">上下文窗口</span>
                  <span class="font-semibold">202K tokens</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">稀疏度</span>
                  <span class="font-semibold">5.9%</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">计算压缩</span>
                  <span class="font-semibold">97%</span>
                </div>
              </div>
            </div>

            <div class="performance-card">
              <div class="flex items-center mb-4">
                <i class="fas fa-trophy text-2xl text-purple-600 mr-3"></i>
                <h3 class="serif text-lg font-semibold">性能表现</h3>
              </div>
              <div class="space-y-3">
                <div class="flex justify-between">
                  <span class="text-gray-600">SWE-bench Verified</span>
                  <span class="font-semibold">77.8%</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">Artificial Analysis</span>
                  <span class="font-semibold">#4 全球</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">开源排名</span>
                  <span class="font-semibold">#1</span>
                </div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <!-- Model Overview -->
      <section id="overview" class="py-16 px-8">
        <div class="max-w-6xl mx-auto">
          <h2 class="serif text-4xl font-bold mb-12">模型概述与核心定位</h2>

          <div id="paradigm-shift" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-6">从&#34;Vibe Coding&#34;到&#34;Agentic Engineering&#34;的范式转变</h3>

            <div class="highlight-box">
              <p class="text-lg font-medium mb-4">
                GLM-5的发布标志着智谱AI在大模型发展战略上的根本性转向——从传统的<strong>&#34;Vibe Coding&#34;（氛围编程）</strong>向<strong>&#34;Agentic Engineering&#34;（智能体工程）</strong>的跃迁<a href="#ref-38" class="citation">[38]</a>
                <a href="#ref-454" class="citation">[454]</a>。
              </p>
            </div>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8 mt-8">
              <div class="bg-red-50 border border-red-200 rounded-lg p-6">
                <h4 class="font-semibold text-red-800 mb-4">
                  <i class="fas fa-times-circle mr-2"></i>Vibe Coding 局限性
                </h4>
                <ul class="space-y-2 text-red-700">
                  <li>• 依赖模型直觉的轻量级开发</li>
                  <li>• 局部代码片段生成</li>
                  <li>• 缺乏系统性理解能力</li>
                  <li>• 无法处理复杂项目端到端</li>
                </ul>
              </div>

              <div class="bg-green-50 border border-green-200 rounded-lg p-6">
                <h4 class="font-semibold text-green-800 mb-4">
                  <i class="fas fa-check-circle mr-2"></i>Agentic Engineering 优势
                </h4>
                <ul class="space-y-2 text-green-700">
                  <li>• 自主规划与多步骤执行</li>
                  <li>• 长期记忆保持</li>
                  <li>• 持续学习与环境适应</li>
                  <li>• 完整软件工程生命周期</li>
                </ul>
              </div>
            </div>

            <p class="mt-8 text-lg leading-relaxed">
              这一范式转变的技术驱动力源于GLM-5在三个维度的突破：<strong>预训练规模的显著扩展</strong>（总参数量从355B提升至744B，预训练数据从23T增至28.5T tokens）为模型提供了更丰富的知识储备；<strong>DeepSeek稀疏注意力机制（DSA）的引入</strong>使得模型在保持长上下文建模能力的同时大幅降低了计算成本；最关键的是<strong>Slime异步强化学习框架与新型智能体RL算法的部署</strong>，使模型能够从复杂、长周期的交互中持续学习优化<a href="#ref-389" class="citation">[389]</a>
              <a href="#ref-413" class="citation">[413]</a>。
            </p>
          </div>

          <div id="open-source-positioning" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-6">开源SOTA定位与市场竞争格局</h3>

            <div class="bg-blue-50 border border-blue-200 rounded-lg p-6 mb-8">
              <h4 class="font-semibold text-blue-800 mb-4">市场定位突破</h4>
              <p class="text-blue-700 mb-4">
                GLM-5在全球排名中位列<strong>第四，在开源模型中排名第一</strong>
                <a href="#ref-12" class="citation">[12]</a>
                <a href="#ref-483" class="citation">[483]</a>，成为首个在综合智能指数上突破50分的开源模型。
              </p>
            </div>

            <!-- Competitive Analysis Table -->
            <div class="overflow-x-auto mt-8">
              <table class="w-full bg-white border border-gray-200 rounded-lg shadow-sm">
                <thead class="bg-gray-50">
                  <tr>
                    <th class="px-6 py-4 text-left font-semibold text-gray-900">维度</th>
                    <th class="px-6 py-4 text-left font-semibold text-gray-900">GLM-5</th>
                    <th class="px-6 py-4 text-left font-semibold text-gray-900">主要竞品</th>
                    <th class="px-6 py-4 text-left font-semibold text-gray-900">差异化特征</th>
                  </tr>
                </thead>
                <tbody class="divide-y divide-gray-200">
                  <tr>
                    <td class="px-6 py-4 font-medium">总参数量</td>
                    <td class="px-6 py-4 text-blue-600 font-semibold">744B</td>
                    <td class="px-6 py-4 text-gray-600">DeepSeek-V3.2: 685B; MiniMax-M2.5: 230B</td>
                    <td class="px-6 py-4 text-green-600">开源最大规模</td>
                  </tr>
                  <tr class="bg-gray-50">
                    <td class="px-6 py-4 font-medium">激活参数</td>
                    <td class="px-6 py-4 text-blue-600 font-semibold">40B/44B</td>
                    <td class="px-6 py-4 text-gray-600">DeepSeek-V3.2: 37B; GPT-4o: ~80B（估计）</td>
                    <td class="px-6 py-4 text-green-600">稀疏效率优化</td>
                  </tr>
                  <tr>
                    <td class="px-6 py-4 font-medium">上下文窗口</td>
                    <td class="px-6 py-4 text-blue-600 font-semibold">202K</td>
                    <td class="px-6 py-4 text-gray-600">DeepSeek-V3.2: 128K; Claude-3.5: 200K</td>
                    <td class="px-6 py-4 text-green-600">长Agent任务支持</td>
                  </tr>
                  <tr class="bg-gray-50">
                    <td class="px-6 py-4 font-medium">开源许可</td>
                    <td class="px-6 py-4 text-blue-600 font-semibold">MIT</td>
                    <td class="px-6 py-4 text-gray-600">Llama: 分层许可; Qwen: 有限商用</td>
                    <td class="px-6 py-4 text-green-600">完全开放</td>
                  </tr>
                  <tr>
                    <td class="px-6 py-4 font-medium">训练算力</td>
                    <td class="px-6 py-4 text-blue-600 font-semibold">华为昇腾</td>
                    <td class="px-6 py-4 text-gray-600">主流: NVIDIA GPU</td>
                    <td class="px-6 py-4 text-green-600">国产自主可控</td>
                  </tr>
                </tbody>
              </table>
            </div>

            <div class="mt-8 bg-yellow-50 border border-yellow-200 rounded-lg p-6">
              <h4 class="font-semibold text-yellow-800 mb-4">
                <i class="fas fa-lightbulb mr-2"></i>开源策略的深远意义
              </h4>
              <p class="text-yellow-700">
                GLM-5采用<strong>MIT许可证</strong>发布模型权重，允许无限制的商用、修改和再分发<a href="#ref-62" class="citation">[62]</a>。这一决策背后有明确的<strong>算力自主化考量</strong>：完全基于华为昇腾芯片和MindSpore框架训练，实现了对国产算力栈的完整验证<a href="#ref-9" class="citation">[9]</a>。
              </p>
            </div>
          </div>

          <!-- Architecture Specifications -->
          <div class="bg-white border border-gray-200 rounded-lg p-8 shadow-sm">
            <h3 class="serif text-2xl font-semibold mb-6">基础规格参数</h3>

            <div class="grid grid-cols-1 md:grid-cols-3 gap-6">
              <div class="text-center">
                <div class="text-3xl font-bold text-blue-600 mb-2">744B</div>
                <div class="text-gray-600">总参数量</div>
                <div class="text-sm text-gray-500 mt-1">激活参数 40B/44B</div>
              </div>
              <div class="text-center">
                <div class="text-3xl font-bold text-green-600 mb-2">28.5T</div>
                <div class="text-gray-600">预训练数据</div>
                <div class="text-sm text-gray-500 mt-1">tokens 规模</div>
              </div>
              <div class="text-center">
                <div class="text-3xl font-bold text-purple-600 mb-2">202K</div>
                <div class="text-gray-600">上下文窗口</div>
                <div class="text-sm text-gray-500 mt-1">最大输出 128K</div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <!-- Architecture Section -->
      <section id="architecture" class="py-16 px-8 bg-slate-50">
        <div class="max-w-6xl mx-auto">
          <h2 class="serif text-4xl font-bold mb-12">模型架构创新</h2>

          <!-- Architecture Diagram -->
          <div class="architecture-diagram mb-12">
            <h3 class="serif text-xl font-semibold mb-6 text-center">GLM-5 架构概览</h3>
            <div class="mermaid-container">
              <div class="mermaid-controls">
                <button class="mermaid-control-btn zoom-in" title="放大">
                  <i class="fas fa-search-plus"></i>
                </button>
                <button class="mermaid-control-btn zoom-out" title="缩小">
                  <i class="fas fa-search-minus"></i>
                </button>
                <button class="mermaid-control-btn reset-zoom" title="重置">
                  <i class="fas fa-expand-arrows-alt"></i>
                </button>
                <button class="mermaid-control-btn fullscreen" title="全屏查看">
                  <i class="fas fa-expand"></i>
                </button>
              </div>
              <div class="mermaid">
                graph TB
                A[&#34;输入序列
                <br/>202K tokens&#34;] --&gt; B[&#34;嵌入层
                <br/>Embedding Layer&#34;]
                B --&gt; C[&#34;前3层稠密FFN
                <br/>Dense FeedForward&#34;]
                C --&gt; D[&#34;MoE层 1-75
                <br/>256 Experts, 8 Active&#34;]
                D --&gt; E[&#34;DSA注意力
                <br/>DeepSeek Sparse Attention&#34;]
                E --&gt; F[&#34;MTP多Token预测
                <br/>Multi-Token Prediction&#34;]
                F --&gt; G[&#34;输出层
                <br/>Output Layer&#34;]

                H[&#34;Lightning Indexer&#34;] --&gt; E
                I[&#34;稀疏选择
                <br/>Top-2048&#34;] --&gt; E

                style A fill:#e1f5fe
                style B fill:#f3e5f5
                style C fill:#e8f5e8
                style D fill:#fff3e0
                style E fill:#fce4ec
                style F fill:#e0f2f1
                style G fill:#f1f8e9
                style H fill:#efebe9
                style I fill:#efebe9
              </div>
            </div>
          </div>

          <div id="moe-architecture" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">混合专家架构（MoE）</h3>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8">
              <div>
                <h4 class="text-lg font-semibold mb-4">256个专家网络设计</h4>
                <p class="mb-4">
                  GLM-5的MoE架构包含<strong>256个专家网络</strong>，每个输入token仅路由至其中最相关的8个专家进行处理，稀疏度约为<strong>5.9%</strong>
                  <a href="#ref-6" class="citation">[6]</a>
                  <a href="#ref-7" class="citation">[7]</a>。
                </p>

                <div class="bg-blue-50 border border-blue-200 rounded-lg p-4 mb-4">
                  <h5 class="font-semibold text-blue-800 mb-2">分层设计策略</h5>
                  <ul class="text-blue-700 space-y-1 text-sm">
                    <li>• 前3层：稠密FFN，确保基础稳定性</li>
                    <li>• 后75层：MoE结构，实现专家specialization</li>
                    <li>• 总计78层隐藏层深度配置</li>
                  </ul>
                </div>
              </div>

              <div>
                <img src="https://kimi-web-img.moonshot.cn/img/www.aitntnews.com/506a582ce66aaa7682f8cb7f08a0faa4cb01f9ce.png" alt="混合专家模型架构示意图" class="w-full h-64 object-cover rounded-lg shadow-md" size="medium" aspect="wide" query="混合专家模型架构" referrerpolicy="no-referrer" data-modified="1" data-score="0.00"/>
              </div>
            </div>
          </div>

          <div id="dsa-mechanism" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">DeepSeek稀疏注意力机制（DSA）</h3>

            <div class="highlight-box">
              <h4 class="text-lg font-semibold mb-4">核心设计目标</h4>
              <p class="mb-4">
                DSA的核心目标是将自注意力机制的计算复杂度从序列长度的平方级<strong>O(L²)降至线性或近线性级别</strong>。对于128K tokens的上下文，全注意力需要计算约82亿个注意力对，而DSA通过选择性稀疏化，将有效计算压缩至约2.6亿对，压缩比达<strong>97%</strong>
                <a href="#ref-34" class="citation">[34]</a>。
              </p>
            </div>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8 mt-8">
              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-blue-800">
                  <i class="fas fa-bolt mr-2"></i>Lightning Indexer
                </h4>
                <ul class="space-y-3 text-gray-700">
                  <li class="flex items-start">
                    <i class="fas fa-check text-green-500 mr-2 mt-1"></i>
                    <span>轻量级评分组件，快速扫描历史token相关性</span>
                  </li>
                  <li class="flex items-start">
                    <i class="fas fa-check text-green-500 mr-2 mt-1"></i>
                    <span>ReLU激活函数替代Softmax，计算效率提升</span>
                  </li>
                  <li class="flex items-start">
                    <i class="fas fa-check text-green-500 mr-2 mt-1"></i>
                    <span>固定Top-k=2048稀疏选择策略</span>
                  </li>
                </ul>
              </div>

              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-green-800">
                  <i class="fas fa-cogs mr-2"></i>两阶段计算流程
                </h4>
                <div class="space-y-4">
                  <div class="border-l-4 border-blue-400 pl-4">
                    <div class="font-medium text-blue-800">阶段一：相关性打分与筛选</div>
                    <div class="text-sm text-gray-600">O(L·d) 复杂度，快速筛选</div>
                  </div>
                  <div class="border-l-4 border-green-400 pl-4">
                    <div class="font-medium text-green-800">阶段二：完整注意力计算</div>
                    <div class="text-sm text-gray-600">O(L·k) 复杂度，k=2048</div>
                  </div>
                </div>
              </div>
            </div>
          </div>

          <div id="mtp-optimization" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">多Token预测（MTP）</h3>

            <div class="bg-purple-50 border border-purple-200 rounded-lg p-6">
              <h4 class="text-lg font-semibold text-purple-800 mb-4">提升生成效率的辅助机制</h4>
              <p class="text-purple-700 mb-4">
                MTP的核心思想是在每个解码步骤中并行预测多个未来token，而非传统的单token自回归生成。这一机制可以显著降低生成延迟，尤其在需要长文本输出的场景中。
              </p>

              <div class="grid grid-cols-1 md:grid-cols-2 gap-4 mt-4">
                <div class="bg-white rounded-lg p-4">
                  <h5 class="font-semibold mb-2">与DSA的协同优化</h5>
                  <ul class="text-sm space-y-1 text-gray-700">
                    <li>• 稀疏注意力降低计算开销</li>
                    <li>• 更多计算预算分配给MTP</li>
                    <li>• 延迟降低补偿两阶段计算开销</li>
                  </ul>
                </div>
                <div class="bg-white rounded-lg p-4">
                  <h5 class="font-semibold mb-2">推测解码策略</h5>
                  <ul class="text-sm space-y-1 text-gray-700">
                    <li>• 每次预测1个额外token</li>
                    <li>• 基于验证的生成机制</li>
                    <li>• 推理框架深度优化</li>
                  </ul>
                </div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <!-- Training Section -->
      <section id="training" class="py-16 px-8">
        <div class="max-w-6xl mx-auto">
          <h2 class="serif text-4xl font-bold mb-12">训练方法与基础设施</h2>

          <div id="pretraining" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">预训练策略</h3>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8">
              <div>
                <h4 class="text-lg font-semibold mb-4">数据规模扩展</h4>
                <p class="mb-4">
                  预训练数据从23T到28.5T的扩展，增幅24%，显著低于参数规模的109%增长。这一&#34;参数增长快于数据增长&#34;的策略，反映了高质量公开文本数据的枯竭挑战<a href="#ref-25" class="citation">[25]</a>
                  <a href="#ref-26" class="citation">[26]</a>。
                </p>

                <div class="bg-orange-50 border border-orange-200 rounded-lg p-4">
                  <h5 class="font-semibold text-orange-800 mb-2">数据策展策略</h5>
                  <ul class="text-orange-700 space-y-1 text-sm">
                    <li>• 更大规模GitHub代码库挖掘</li>
                    <li>• 代码-文档-提交历史联合建模</li>
                    <li>• 合成代码数据生成与筛选</li>
                    <li>• 多语言比例精心调配</li>
                  </ul>
                </div>
              </div>

              <div>
                <img src="https://kimi-web-img.moonshot.cn/img/event.asus.com.cn/e62a0ceeeb9738bb3171232cfbbbb643d58ee078.jpg" alt="AI训练数据中心的机架式服务器集群" class="w-full h-64 object-cover rounded-lg shadow-md" size="medium" aspect="wide" style="photo" query="AI训练数据中心服务器集群" referrerpolicy="no-referrer" data-modified="1" data-score="0.00"/>
              </div>
            </div>
          </div>

          <div id="slime-framework" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">Slime异步强化学习框架</h3>

            <div class="highlight-box">
              <h4 class="text-lg font-semibold mb-4">核心创新：生成与训练的解耦架构</h4>
              <p class="mb-4">
                Slime是GLM-5训练方法中最具原创性的技术贡献，被官方描述为&#34;新型异步强化学习基础设施&#34;<a href="#ref-9" class="citation">[9]</a>
                <a href="#ref-30" class="citation">[30]</a>。其命名&#34;Slime&#34;（史莱姆）暗示了系统的灵活性与适应性。
              </p>
            </div>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8 mt-8">
              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-blue-800">
                  <i class="fas fa-sync-alt mr-2"></i>异步架构优势
                </h4>
                <ul class="space-y-3 text-gray-700">
                  <li class="flex items-start">
                    <i class="fas fa-arrow-up text-green-500 mr-2 mt-1"></i>
                    <span>GPU利用率接近100%，消除同步等待</span>
                  </li>
                  <li class="flex items-start">
                    <i class="fas fa-shield-alt text-blue-500 mr-2 mt-1"></i>
                    <span>重要性采样校正，避免分布偏移问题</span>
                  </li>
                  <li class="flex items-start">
                    <i class="fas fa-tachometer-alt text-purple-500 mr-2 mt-1"></i>
                    <span>支持更大模型规模及更复杂RL任务</span>
                  </li>
                </ul>
              </div>

              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-green-800">
                  <i class="fas fa-brain mr-2"></i>智能体RL算法
                </h4>
                <ul class="space-y-3 text-gray-700">
                  <li class="flex items-start">
                    <i class="fas fa-sitemap text-blue-500 mr-2 mt-1"></i>
                    <span>长程任务自动分解机制</span>
                  </li>
                  <li class="flex items-start">
                    <i class="fas fa-crosshairs text-green-500 mr-2 mt-1"></i>
                    <span>跨步骤奖励归因与信用分配</span>
                  </li>
                  <li class="flex items-start">
                    <i class="fas fa-redo text-orange-500 mr-2 mt-1"></i>
                    <span>轨迹回放与延迟奖励分配</span>
                  </li>
                </ul>
              </div>
            </div>

            <!-- Slime Architecture Diagram -->
            <div class="architecture-diagram mt-8">
              <h4 class="text-lg font-semibold mb-6 text-center">Slime异步RL架构</h4>
              <div class="mermaid-container">
                <div class="mermaid-controls">
                  <button class="mermaid-control-btn zoom-in" title="放大">
                    <i class="fas fa-search-plus"></i>
                  </button>
                  <button class="mermaid-control-btn zoom-out" title="缩小">
                    <i class="fas fa-search-minus"></i>
                  </button>
                  <button class="mermaid-control-btn reset-zoom" title="重置">
                    <i class="fas fa-expand-arrows-alt"></i>
                  </button>
                  <button class="mermaid-control-btn fullscreen" title="全屏查看">
                    <i class="fas fa-expand"></i>
                  </button>
                </div>
                <div class="mermaid">
                  graph LR
                  A[&#34;Rollout Worker Pool&#34;] --&gt; B[&#34;Experience Buffer&#34;]
                  C[&#34;Training Worker Pool&#34;] --&gt; B
                  B --&gt; D[&#34;Gradient Updates&#34;]
                  D --&gt; A
                  D --&gt; C

                  E[&#34;Environment&#34;] --&gt; A
                  F[&#34;Reward Model&#34;] --&gt; B

                  G[&#34;Offline Data&#34;] --&gt; B
                  H[&#34;Human Feedback&#34;] --&gt; B

                  style A fill:#e1f5fe
                  style B fill:#f3e5f5
                  style C fill:#e8f5e8
                  style D fill:#fff3e0
                  style E fill:#fce4ec
                  style F fill:#e0f2f1
                  style G fill:#f1f8e9
                  style H fill:#efebe9
                </div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <!-- Performance Section -->
      <section id="performance" class="py-16 px-8 bg-slate-50">
        <div class="max-w-6xl mx-auto">
          <h2 class="serif text-4xl font-bold mb-12">性能表现与基准测试</h2>

          <div id="academic-benchmarks" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">学术基准测试</h3>

            <div class="grid grid-cols-1 lg:grid-cols-3 gap-6 mt-8">
              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-blue-800">
                  <i class="fas fa-code mr-2"></i>编程能力
                </h4>
                <div class="space-y-3">
                  <div class="flex justify-between items-center">
                    <span class="text-gray-600">SWE-bench Verified</span>
                    <span class="text-xl font-bold text-blue-600">77.8%</span>
                  </div>
                  <div class="flex justify-between items-center">
                    <span class="text-gray-600">Terminal Bench 2.0</span>
                    <span class="text-xl font-bold text-blue-600">56.2</span>
                  </div>
                  <div class="text-xs text-gray-500 mt-2">
                    开源第一，接近Claude Opus 4.5
                  </div>
                </div>
              </div>

              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-green-800">
                  <i class="fas fa-tools mr-2"></i>工具推理
                </h4>
                <div class="space-y-3">
                  <div class="flex justify-between items-center">
                    <span class="text-gray-600">Humanity&#39;s Last Exam</span>
                    <span class="text-xl font-bold text-green-600">50.4%</span>
                  </div>
                  <div class="flex justify-between items-center">
                    <span class="text-gray-600">BrowseComp</span>
                    <span class="text-xl font-bold text-green-600">领先</span>
                  </div>
                  <div class="text-xs text-gray-500 mt-2">
                    工具使用与推理能力
                  </div>
                </div>
              </div>

              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-purple-800">
                  <i class="fas fa-robot mr-2"></i>Agent任务
                </h4>
                <div class="space-y-3">
                  <div class="flex justify-between items-center">
                    <span class="text-gray-600">Vending Bench 2</span>
                    <span class="text-xl font-bold text-purple-600">$4,432</span>
                  </div>
                  <div class="flex justify-between items-center">
                    <span class="text-gray-600">MCP-Atlas</span>
                    <span class="text-xl font-bold text-purple-600">领先</span>
                  </div>
                  <div class="text-xs text-gray-500 mt-2">
                    长程决策与执行能力
                  </div>
                </div>
              </div>
            </div>

            <!-- Performance Comparison Table -->
            <div class="mt-12 bg-white border border-gray-200 rounded-lg p-6 shadow-sm">
              <h4 class="text-lg font-semibold mb-6">GLM-5 基准测试成绩汇总</h4>
              <div class="overflow-x-auto">
                <table class="w-full">
                  <thead class="bg-gray-50">
                    <tr>
                      <th class="px-4 py-3 text-left font-semibold">基准测试</th>
                      <th class="px-4 py-3 text-left font-semibold">GLM-5得分</th>
                      <th class="px-4 py-3 text-left font-semibold">对比基准</th>
                      <th class="px-4 py-3 text-left font-semibold">排名</th>
                    </tr>
                  </thead>
                  <tbody class="divide-y divide-gray-200">
                    <tr class="bg-blue-50">
                      <td class="px-4 py-3 font-medium">SWE-bench Verified</td>
                      <td class="px-4 py-3 text-blue-600 font-bold">77.8</td>
                      <td class="px-4 py-3 text-sm">Claude Opus 4.5 ~79</td>
                      <td class="px-4 py-3">
                        <span class="px-2 py-1 bg-blue-100 text-blue-800 rounded text-xs font-medium">开源第一</span>
                      </td>
                    </tr>
                    <tr>
                      <td class="px-4 py-3 font-medium">Terminal Bench 2.0</td>
                      <td class="px-4 py-3 text-green-600 font-bold">56.2</td>
                      <td class="px-4 py-3 text-sm">Claude Opus 4.5 ~58</td>
                      <td class="px-4 py-3">
                        <span class="px-2 py-1 bg-green-100 text-green-800 rounded text-xs font-medium">开源第一</span>
                      </td>
                    </tr>
                    <tr class="bg-gray-50">
                      <td class="px-4 py-3 font-medium">Humanity&#39;s Last Exam</td>
                      <td class="px-4 py-3 text-purple-600 font-bold">50.4%</td>
                      <td class="px-4 py-3 text-sm">GLM-4.7 42.8%</td>
                      <td class="px-4 py-3">
                        <span class="px-2 py-1 bg-purple-100 text-purple-800 rounded text-xs font-medium">开源第一</span>
                      </td>
                    </tr>
                    <tr>
                      <td class="px-4 py-3 font-medium">BrowseComp</td>
                      <td class="px-4 py-3 text-orange-600 font-bold">领先</td>
                      <td class="px-4 py-3 text-sm">Gemini 3 Pro等</td>
                      <td class="px-4 py-3">
                        <span class="px-2 py-1 bg-orange-100 text-orange-800 rounded text-xs font-medium">开源第一</span>
                      </td>
                    </tr>
                    <tr class="bg-gray-50">
                      <td class="px-4 py-3 font-medium">Vending Bench 2</td>
                      <td class="px-4 py-3 text-red-600 font-bold">$4,432</td>
                      <td class="px-4 py-3 text-sm">Claude Opus 4.5 ~$4,500</td>
                      <td class="px-4 py-3">
                        <span class="px-2 py-1 bg-red-100 text-red-800 rounded text-xs font-medium">开源第一</span>
                      </td>
                    </tr>
                  </tbody>
                </table>
              </div>
              <div class="text-xs text-gray-500 mt-4">
                数据来源：<a href="#ref-1" class="citation">[1]</a>
                <a href="#ref-3" class="citation">[3]</a>
                <a href="#ref-12" class="citation">[12]</a>
                <a href="#ref-26" class="citation">[26]</a>
                <a href="#ref-30" class="citation">[30]</a>
              </div>
            </div>
          </div>

          <div id="real-world" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">真实场景能力</h3>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8">
              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-blue-800">
                  <i class="fas fa-project-diagram mr-2"></i>端到端软件工程
                </h4>
                <p class="mb-4 text-gray-700">
                  从需求文档自动生成可部署的微服务、配套测试与CI配置<a href="#ref-21" class="citation">[21]</a>。开发完整的横版解谜游戏、Agent交互世界等应用<a href="#ref-47" class="citation">[47]</a>。
                </p>
                <div class="bg-blue-50 border border-blue-200 rounded-lg p-3">
                  <div class="text-sm text-blue-800">
                    <strong>CC-Bench-V2评估：</strong>较GLM-4.7平均提升超过20%
                  </div>
                </div>
              </div>

              <div class="performance-card">
                <h4 class="text-lg font-semibold mb-4 text-green-800">
                  <i class="fas fa-cogs mr-2"></i>复杂系统工程
                </h4>
                <ul class="space-y-2 text-gray-700">
                  <li>• Mac系统界面模拟实现<a href="#ref-10" class="citation">[10]</a>
                  </li>
                  <li>• GBA模拟器完整开发<a href="#ref-17" class="citation">[17]</a>
                  </li>
                  <li>• 3D渲染与游戏逻辑</li>
                  <li>• 跨组件协调与架构设计</li>
                </ul>
              </div>
            </div>

            <div class="mt-8 bg-green-50 border border-green-200 rounded-lg p-6">
              <h4 class="font-semibold text-green-800 mb-4">
                <i class="fas fa-chart-line mr-2"></i>Vending Bench 2 表现
              </h4>
              <p class="text-green-700 mb-4">
                在模拟的一年时间跨度内经营自动售货机业务，涉及库存管理、定价策略、需求预测、财务决策。GLM-5的最终账户余额<strong>$4,432</strong>，接近Claude Opus 4.5的$4,500<a href="#ref-12" class="citation">[12]</a>
                <a href="#ref-30" class="citation">[30]</a>。
              </p>
            </div>
          </div>
        </div>
      </section>

      <!-- Deployment Section -->
      <section id="deployment" class="py-16 px-8">
        <div class="max-w-6xl mx-auto">
          <h2 class="serif text-4xl font-bold mb-12">工程实现与部署优化</h2>

          <div id="token-efficiency" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">Token效率优化</h3>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8">
              <div>
                <h4 class="text-lg font-semibold mb-4">DSA实际收益量化</h4>
                <div class="space-y-4">
                  <div class="bg-blue-50 border border-blue-200 rounded-lg p-4">
                    <div class="flex justify-between items-center mb-2">
                      <span class="font-semibold text-blue-800">理论计算压缩</span>
                      <span class="text-xl font-bold text-blue-600">97%</span>
                    </div>
                    <div class="text-sm text-blue-700">
                      从O(L²/2)≈8.2×10⁹降至O(L·k)=2.6×10⁸
                    </div>
                  </div>

                  <div class="bg-green-50 border border-green-200 rounded-lg p-4">
                    <div class="flex justify-between items-center mb-2">
                      <span class="font-semibold text-green-800">内存占用优化</span>
                      <span class="text-xl font-bold text-green-600">显著降低</span>
                    </div>
                    <div class="text-sm text-green-700">
                      注意力权重存储需求大幅下降
                    </div>
                  </div>
                </div>
              </div>

              <div>
                <img src="https://kimi-web-img.moonshot.cn/img/preview.redd.it/1c02e0b3902e38b7a1edbf3fd554148748a731ca.png" alt="服务器机架中的GPU集群" class="w-full h-64 object-cover rounded-lg shadow-md" size="medium" aspect="wide" style="photo" query="服务器GPU集群" referrerpolicy="no-referrer" data-modified="1" data-score="0.00"/>
              </div>
            </div>
          </div>

          <div id="hardware" class="mb-16">
            <h3 class="serif text-2xl font-semibold mb-8">硬件适配与可扩展性</h3>

            <div class="highlight-box">
              <h4 class="text-lg font-semibold mb-4">全国产化战略</h4>
              <p class="mb-4">
                GLM-5的硬件适配策略体现了&#34;全国产化&#34;的战略导向。技术报告披露的适配平台涵盖了华为昇腾、摩尔线程、寒武纪、昆仑芯、沐曦、燧原、海光等主要国产AI芯片厂商<a href="#ref-1" class="citation">[1]</a>
                <a href="#ref-2" class="citation">[2]</a>。
              </p>
            </div>

            <div class="grid grid-cols-1 md:grid-cols-2 lg:grid-cols-4 gap-6 mt-8">
              <div class="performance-card text-center">
                <i class="fas fa-server text-3xl text-blue-600 mb-3"></i>
                <h4 class="font-semibold mb-2">华为昇腾</h4>
                <div class="text-sm text-gray-600">Ascend</div>
              </div>
              <div class="performance-card text-center">
                <i class="fas fa-microchip text-3xl text-green-600 mb-3"></i>
                <h4 class="font-semibold mb-2">摩尔线程</h4>
                <div class="text-sm text-gray-600">Moore Threads</div>
              </div>
              <div class="performance-card text-center">
                <i class="fas fa-brain text-3xl text-purple-600 mb-3"></i>
                <h4 class="font-semibold mb-2">寒武纪</h4>
                <div class="text-sm text-gray-600">Cambricon</div>
              </div>
              <div class="performance-card text-center">
                <i class="fas fa-network-wired text-3xl text-orange-600 mb-3"></i>
                <h4 class="font-semibold mb-2">昆仑芯</h4>
                <div class="text-sm text-gray-600">Kunlun</div>
              </div>
            </div>

            <div class="mt-8 bg-yellow-50 border border-yellow-200 rounded-lg p-6">
              <h4 class="font-semibold text-yellow-800 mb-4">
                <i class="fas fa-dollar-sign mr-2"></i>部署成本优势
              </h4>
              <p class="text-yellow-700">
                通过这些优化，GLM-5在国产芯片集群上实现了&#34;高吞吐、低延迟的稳定运行&#34;，与双GPU国际集群的部署成本相比<strong>&#34;减半&#34;</strong>
                <a href="#ref-51" class="citation">[51]</a>。
              </p>
            </div>
          </div>

          <!-- Open Source Ecosystem -->
          <div class="bg-white border border-gray-200 rounded-lg p-8 shadow-sm">
            <h3 class="serif text-2xl font-semibold mb-6">开源生态建设</h3>

            <div class="grid grid-cols-1 lg:grid-cols-2 gap-8">
              <div>
                <h4 class="text-lg font-semibold mb-4">技术透明化</h4>
                <div class="space-y-3">
                  <div class="flex items-center">
                    <i class="fas fa-check-circle text-green-500 mr-3"></i>
                    <span>模型权重（MIT许可证）完全开放</span>
                  </div>
                  <div class="flex items-center">
                    <i class="fas fa-check-circle text-green-500 mr-3"></i>
                    <span>训练代码与Slime框架开源</span>
                  </div>
                  <div class="flex items-center">
                    <i class="fas fa-check-circle text-green-500 mr-3"></i>
                    <span>训练日志详细记录</span>
                  </div>
                  <div class="flex items-center">
                    <i class="fas fa-check-circle text-green-500 mr-3"></i>
                    <span>社区驱动的技术透明化</span>
                  </div>
                </div>
              </div>

              <div>
                <h4 class="text-lg font-semibold mb-4">社区参与</h4>
                <div class="bg-blue-50 border border-blue-200 rounded-lg p-4">
                  <h5 class="font-semibold text-blue-800 mb-2">Pony Alpha 匿名测试</h5>
                  <p class="text-blue-700 text-sm">
                    发布前通过匿名身份在OpenRouter平台测试，获得社区91%以上用户对其身份的准确判断，并登顶热度榜首<a href="#ref-6" class="citation">[6]</a>
                    <a href="#ref-24" class="citation">[24]</a>。
                  </p>
                </div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <!-- Evolution Section -->
      <section id="evolution" class="py-16 px-8 bg-slate-50">
        <div class="max-w-6xl mx-auto">
          <h2 class="serif text-4xl font-bold mb-12">技术演进与架构溯源</h2>

          <div class="grid grid-cols-1 lg:grid-cols-2 gap-8 mb-16">
            <div class="performance-card">
              <h3 class="serif text-xl font-semibold mb-4 text-blue-800">
                <i class="fas fa-code-branch mr-2"></i>DeepSeek-V3/V3.2架构继承
              </h3>
              <p class="mb-4 text-gray-700">
                GLM-5对DeepSeek-V3/V3.2架构的继承，是理解其技术路线选择的关键。代码审查确认，GLM-5的DSA实现直接继承自DeepSeek的代码库<a href="#ref-6" class="citation">[6]</a>
                <a href="#ref-34" class="citation">[34]</a>。
              </p>
              <div class="bg-blue-50 border border-blue-200 rounded-lg p-3">
                <div class="text-sm text-blue-800">
                  <strong>策略合理性：</strong>降低研发风险、加速产品迭代、复用成熟优化
                </div>
              </div>
            </div>

            <div class="performance-card">
              <h3 class="serif text-xl font-semibold mb-4 text-green-800">
                <i class="fas fa-arrow-up mr-2"></i>相对GLM-4.7的架构升级
              </h3>
              <div class="space-y-3">
                <div class="flex justify-between">
                  <span class="text-gray-600">总参数</span>
                  <span class="font-semibold text-green-600">355B → 744B (+109%)</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">激活参数</span>
                  <span class="font-semibold text-blue-600">32B → 40B (+25%)</span>
                </div>
                <div class="flex justify-between">
                  <span class="text-gray-600">上下文窗口</span>
                  <span class="font-semibold text-purple-600">扩展至202K</span>
                </div>
              </div>
            </div>
          </div>

          <!-- Architecture Evolution Diagram -->
          <div class="architecture-diagram mb-16">
            <h3 class="serif text-xl font-semibold mb-6 text-center">GLM系列技术演进路径</h3>
            <div class="mermaid-container">
              <div class="mermaid-controls">
                <button class="mermaid-control-btn zoom-in" title="放大">
                  <i class="fas fa-search-plus"></i>
                </button>
                <button class="mermaid-control-btn zoom-out" title="缩小">
                  <i class="fas fa-search-minus"></i>
                </button>
                <button class="mermaid-control-btn reset-zoom" title="重置">
                  <i class="fas fa-expand-arrows-alt"></i>
                </button>
                <button class="mermaid-control-btn fullscreen" title="全屏查看">
                  <i class="fas fa-expand"></i>
                </button>
              </div>
              <div class="mermaid">
                graph TD
                A[&#34;GLM-4.7
                <br/>355B参数&#34;] --&gt; B[&#34;技术积累&#34;]
                B --&gt; C[&#34;GLM-5
                <br/>744B参数&#34;]
                A --&gt; D[&#34;MoE架构&#34;]
                A --&gt; E[&#34;代码能力&#34;]
                A --&gt; F[&#34;长上下文&#34;]

                D --&gt; G[&#34;256专家MoE
                <br/>稀疏度5.9%&#34;]
                E --&gt; H[&#34;Slime异步RL
                <br/>Agentic Engineering&#34;]
                F --&gt; I[&#34;DSA注意力
                <br/>202K上下文&#34;]

                J[&#34;DeepSeek-V3.2&#34;] --&gt; K[&#34;DSA技术迁移&#34;]
                K --&gt; I

                C --&gt; L[&#34;开源SOTA
                <br/>全球#4&#34;]

                style A fill:#e1f5fe
                style B fill:#f3e5f5
                style C fill:#e8f5e8
                style D fill:#fff3e0
                style E fill:#fce4ec
                style F fill:#e0f2f1
                style G fill:#f1f8e9
                style H fill:#efebe9
                style I fill:#e8eaf6
                style J fill:#fff8e1
                style K fill:#f3e5f5
                style L fill:#e8f5e8
              </div>
            </div>
          </div>

          <div class="bg-white border border-gray-200 rounded-lg p-8 shadow-sm">
            <h3 class="serif text-xl font-semibold mb-6">社区驱动的技术透明化</h3>
            <p class="mb-4 text-gray-700">
              GLM-5的技术信息披露模式具有鲜明的&#34;社区驱动&#34;特征。详细的架构参数、部署配置、性能数据，很大程度上来自社区的分析与挖掘，而非官方的系统性发布<a href="#ref-18" class="citation">[18]</a>
              <a href="#ref-40" class="citation">[40]</a>。
            </p>

            <div class="grid grid-cols-1 md:grid-cols-2 gap-6">
              <div class="bg-green-50 border border-green-200 rounded-lg p-4">
                <h4 class="font-semibold text-green-800 mb-2">优势</h4>
                <ul class="text-green-700 space-y-1 text-sm">
                  <li>• 激发社区技术参与热情</li>
                  <li>• 加速问题发现与解决</li>
                  <li>• 促进技术生态建设</li>
                </ul>
              </div>
              <div class="bg-orange-50 border border-orange-200 rounded-lg p-4">
                <h4 class="font-semibold text-orange-800 mb-2">挑战</h4>
                <ul class="text-orange-700 space-y-1 text-sm">
                  <li>• 信息碎片化与不准确性</li>
                  <li>• 关键细节披露不完整</li>
                  <li>• 依赖社区分析能力</li>
                </ul>
              </div>
            </div>
          </div>
        </div>
      </section>

      <!-- Limitations and Future -->
      <section id="limitations" class="py-16 px-8">
        <div class="max-w-6xl mx-auto">
          <h2 class="serif text-4xl font-bold mb-12">研究局限与未来方向</h2>

          <div class="bg-red-50 border border-red-200 rounded-lg p-6 mb-12">
            <h3 class="serif text-xl font-semibold mb-4 text-red-800">
              <i class="fas fa-exclamation-triangle mr-2"></i>官方技术细节的披露程度
            </h3>
            <p class="text-red-700 mb-4">
              当前对GLM-5的技术理解受限于信息披露的不完整性。关键未公开细节包括DSA中Lightning Indexer的网络结构、Slime框架的异步通信协议、预训练数据的详细构成等。
            </p>
          </div>

          <div class="grid grid-cols-1 lg:grid-cols-2 gap-8 mb-16">
            <div class="performance-card">
              <h3 class="serif text-xl font-semibold mb-6 text-blue-800">
                <i class="fas fa-rocket mr-2"></i>技术突破方向
              </h3>
              <div class="space-y-4">
                <div class="border-l-4 border-blue-400 pl-4">
                  <h4 class="font-semibold text-blue-800">超长上下文泛化</h4>
                  <p class="text-sm text-gray-600 mt-1">固定k=2048的瓶颈突破，层次化索引设计</p>
                </div>
                <div class="border-l-4 border-green-400 pl-4">
                  <h4 class="font-semibold text-green-800">多智能体协同</h4>
                  <p class="text-sm text-gray-600 mt-1">角色分工、通信协议、群体智能系统</p>
                </div>
                <div class="border-l-4 border-purple-400 pl-4">
                  <h4 class="font-semibold text-purple-800">持续学习适应</h4>
                  <p class="text-sm text-gray-600 mt-1">灾难性遗忘缓解，终身学习机制</p>
                </div>
              </div>
            </div>

            <div class="performance-card">
              <h3 class="serif text-xl font-semibold mb-6 text-green-800">
                <i class="fas fa-lightbulb mr-2"></i>应用场景拓展
              </h3>
              <div class="space-y-3">
                <div class="flex items-center">
                  <i class="fas fa-code text-blue-500 mr-3"></i>
                  <span>企业级软件开发</span>
                </div>
                <div class="flex items-center">
                  <i class="fas fa-cog text-green-500 mr-3"></i>
                  <span>自动化系统工程</span>
                </div>
                <div class="flex items-center">
                  <i class="fas fa-graduation-cap text-purple-500 mr-3"></i>
                  <span>教育辅助系统</span>
                </div>
                <div class="flex items-center">
                  <i class="fas fa-chart-line text-orange-500 mr-3"></i>
                  <span>科研辅助分析</span>
                </div>
              </div>
            </div>
          </div>

          <!-- Future Research Directions Table -->
          <div class="bg-white border border-gray-200 rounded-lg p-8 shadow-sm">
            <h3 class="serif text-2xl font-semibold mb-8 text-center">GLM-5 未来研究方向展望</h3>

            <div class="overflow-x-auto">
              <table class="w-full">
                <thead class="bg-gray-50">
                  <tr>
                    <th class="px-6 py-4 text-left font-semibold">研究方向</th>
                    <th class="px-6 py-4 text-left font-semibold">技术挑战</th>
                    <th class="px-6 py-4 text-left font-semibold">潜在突破</th>
                  </tr>
                </thead>
                <tbody class="divide-y divide-gray-200">
                  <tr>
                    <td class="px-6 py-4 font-medium">超长上下文（&gt;1M tokens）DSA泛化</td>
                    <td class="px-6 py-4 text-sm">固定k=2048的瓶颈、层次化索引设计</td>
                    <td class="px-6 py-4 text-sm">自适应稀疏策略、递归注意力</td>
                  </tr>
                  <tr class="bg-gray-50">
                    <td class="px-6 py-4 font-medium">多智能体协同工程</td>
                    <td class="px-6 py-4 text-sm">角色分工、通信协议、冲突解决</td>
                    <td class="px-6 py-4 text-sm">分布式Agent系统、群体智能</td>
                  </tr>
                  <tr>
                    <td class="px-6 py-4 font-medium">持续学习与终身适应</td>
                    <td class="px-6 py-4 text-sm">灾难性遗忘、稳定性-可塑性权衡</td>
                    <td class="px-6 py-4 text-sm">模块化架构、元学习、神经可塑性</td>
                  </tr>
                  <tr class="bg-gray-50">
                    <td class="px-6 py-4 font-medium">多模态能力的深度整合</td>
                    <td class="px-6 py-4 text-sm">跨模态对齐、统一表征空间</td>
                    <td class="px-6 py-4 text-sm">原生多模态架构、端到端训练</td>
                  </tr>
                </tbody>
              </table>
            </div>
          </div>

          <div class="mt-12 text-center">
            <div class="inline-block bg-gradient-to-r from-blue-600 to-purple-600 text-white px-8 py-6 rounded-2xl shadow-lg">
              <h3 class="serif text-2xl font-bold mb-2">GLM-5 里程碑意义</h3>
              <p class="text-lg opacity-90">
                标志着开源大模型在&#34;Agentic Engineering&#34;领域的重要里程碑，为后续研究提供有价值的参考
              </p>
            </div>
          </div>
        </div>
      </section>

      <!-- References Footer -->
      <footer class="py-12 px-8 bg-gray-900 text-white">
        <div class="max-w-6xl mx-auto">
          <h3 class="serif text-2xl font-bold mb-8">参考文献</h3>
          <div class="grid grid-cols-1 md:grid-cols-2 lg:grid-cols-3 gap-4 text-sm">
            <div class="space-y-2">
              <div id="ref-1">[1] <a href="https://m.36kr.com/p/3679762791182213" class="text-blue-400 hover:text-blue-300">36氪 - GLM-5技术报告</a>
              </div>
              <div id="ref-2">[2] <a href="https://www.ifanr.com/1655048" class="text-blue-400 hover:text-blue-300">爱范儿 - AI芯片产业报道</a>
              </div>
              <div id="ref-3">[3] <a href="https://zhuanlan.zhihu.com/p/2006857967409059074" class="text-blue-400 hover:text-blue-300">知乎专栏 - GLM-5分析</a>
              </div>
              <div id="ref-4">[4] <a href="https://learnku.com/articles/91817" class="text-blue-400 hover:text-blue-300">LearnKu - 多模态架构</a>
              </div>
              <div id="ref-5">[5] <a href="https://ascendai.csdn.net/699443c20a2f6a37c59233d4.html" class="text-blue-400 hover:text-blue-300">CSDN - 昇腾AI</a>
              </div>
              <div id="ref-6">[6] <a href="https://hub.baai.ac.cn/view/52489" class="text-blue-400 hover:text-blue-300">BAAI Hub - 架构分析</a>
              </div>
              <div id="ref-7">[7] <a href="https://m.datalearner.com/ai-models/pretrained-models/glm-5" class="text-blue-400 hover:text-blue-300">DataLearner - 模型规格</a>
              </div>
              <div id="ref-9">[9] <a href="https://help.apiyi.com/glm-5-api-guide-744b-moe-agent-tutorial.html" class="text-blue-400 hover:text-blue-300">API指南 - GLM-5教程</a>
              </div>
              <div id="ref-10">[10] <a href="https://www.huxiu.com/article/4834610.html" class="text-blue-400 hover:text-blue-300">虎嗅 - 系统界面实现</a>
              </div>
              <div id="ref-11">[11] <a href="https://milvus.io/zh/blog/glm5-vs-minimax-m25-vs-gemini-3-deep-think.md" class="text-blue-400 hover:text-blue-300">Milvus - 模型对比</a>
              </div>
              <div id="ref-12">[12] <a href="https://hub.baai.ac.cn/view/52524" class="text-blue-400 hover:text-blue-300">BAAI Hub - 性能基准</a>
              </div>
            </div>
            <div class="space-y-2">
              <div id="ref-17">[17] <a href="https://docs.feishu.cn/article/wiki/FjiOwWp2giA7hRk6jjfcPioCnAc" class="text-blue-400 hover:text-blue-300">飞书文档 - GBA模拟器</a>
              </div>
              <div id="ref-18">[18] <a href="https://modelscope.cn/models/ZhipuAI/GLM-5" class="text-blue-400 hover:text-blue-300">ModelScope - GLM-5模型</a>
              </div>
              <div id="ref-21">[21] <a href="https://hub.baai.ac.cn/paper/96b2bf85-fcae-4e27-ac4c-b90022b60d81" class="text-blue-400 hover:text-blue-300">BAAI Hub - 软件工程</a>
              </div>
              <div id="ref-24">[24] <a href="https://companies.caixin.com/m/2026-02-12/102413864.html" class="text-blue-400 hover:text-blue-300">财新网 - Pony Alpha测试</a>
              </div>
              <div id="ref-25">[25] <a href="https://wallstreetcn.com/articles/3765532" class="text-blue-400 hover:text-blue-300">华尔街见闻 - 技术报告</a>
              </div>
              <div id="ref-26">[26] <a href="https://www.paratera.com/news_des/157.html" class="text-blue-400 hover:text-blue-300">并行科技 - 性能评估</a>
              </div>
              <div id="ref-30">[30] <a href="https://z.ai/blog/glm-5" class="text-blue-400 hover:text-blue-300">Z.ai - 官方博客</a>
              </div>
              <div id="ref-34">[34] <a href="https://m.36kr.com/p/3678132374709128" class="text-blue-400 hover:text-blue-300">36氪 - DSA机制分析</a>
              </div>
              <div id="ref-38">[38] <a href="https://github.com/zai-org/GLM-5" class="text-blue-400 hover:text-blue-300">GitHub - GLM-5代码</a>
              </div>
              <div id="ref-40">[40] <a href="https://github.com/zai-org/GLM-5" class="text-blue-400 hover:text-blue-300">GitHub - 技术文档</a>
              </div>
              <div id="ref-47">[47] <a href="https://zhuanlan.zhihu.com/p/2005229751980275261" class="text-blue-400 hover:text-blue-300">知乎专栏 - 应用案例</a>
              </div>
            </div>
            <div class="space-y-2">
              <div id="ref-51">[51] <a href="https://www.emergentmind.com/papers/2602.15763" class="text-blue-400 hover:text-blue-300">Emergent Mind - 部署成本分析</a>
              </div>
              <div id="ref-62">[62] <a href="https://advenboost.com/glm-5-how-to-use-the-worlds-first-50-score-open-source-ai-2026/" class="text-blue-400 hover:text-blue-300">Advenboost - 使用指南</a>
              </div>
              <div id="ref-389">[389] <a href="https://www.emergentmind.com/papers/2602.15763" class="text-blue-400 hover:text-blue-300">Emergent Mind - 技术论文</a>
              </div>
              <div id="ref-413">[413] <a href="https://build.nvidia.com/z-ai/glm5/modelcard" class="text-blue-400 hover:text-blue-300">NVIDIA - 模型卡片</a>
              </div>
              <div id="ref-454">[454] <a href="https://www.arxiv.org/abs/2602.15763" class="text-blue-400 hover:text-blue-300">arXiv - 技术报告</a>
              </div>
              <div id="ref-483">[483] <a href="https://www.jinguxun.com/article/7579033" class="text-blue-400 hover:text-blue-300">金股讯 - 市场分析</a>
              </div>
            </div>
          </div>
        </div>
      </footer>
    </main>

    <script>
        // Initialize Mermaid
        mermaid.initialize({
            startOnLoad: true,
            theme: 'default',
            themeVariables: {
                primaryColor: '#2563eb',
                primaryTextColor: '#1f2937',
                primaryBorderColor: '#374151',
                lineColor: '#6b7280',
                secondaryColor: '#f8fafc',
                tertiaryColor: '#f1f5f9',
                background: '#ffffff',
                mainBkg: '#ffffff',
                secondBkg: '#f8fafc',
                tertiaryBkg: '#f1f5f9'
            },
            flowchart: {
                htmlLabels: true,
                curve: 'basis'
            },
            fontSize: 14,
            fontFamily: 'Inter, sans-serif'
        });

        // Initialize Mermaid Controls for zoom and pan
        function initializeMermaidControls() {
            const containers = document.querySelectorAll('.mermaid-container');

            containers.forEach(container => {
            const mermaidElement = container.querySelector('.mermaid');
            let scale = 1;
            let isDragging = false;
            let startX, startY, translateX = 0, translateY = 0;

            // 触摸相关状态
            let isTouch = false;
            let touchStartTime = 0;
            let initialDistance = 0;
            let initialScale = 1;
            let isPinching = false;

            // Zoom controls
            const zoomInBtn = container.querySelector('.zoom-in');
            const zoomOutBtn = container.querySelector('.zoom-out');
            const resetBtn = container.querySelector('.reset-zoom');
            const fullscreenBtn = container.querySelector('.fullscreen');

            function updateTransform() {
                mermaidElement.style.transform = `translate(${translateX}px, ${translateY}px) scale(${scale})`;

                if (scale > 1) {
                container.classList.add('zoomed');
                } else {
                container.classList.remove('zoomed');
                }

                mermaidElement.style.cursor = isDragging ? 'grabbing' : 'grab';
            }

            if (zoomInBtn) {
                zoomInBtn.addEventListener('click', () => {
                scale = Math.min(scale * 1.25, 4);
                updateTransform();
                });
            }

            if (zoomOutBtn) {
                zoomOutBtn.addEventListener('click', () => {
                scale = Math.max(scale / 1.25, 0.3);
                if (scale <= 1) {
                    translateX = 0;
                    translateY = 0;
                }
                updateTransform();
                });
            }

            if (resetBtn) {
                resetBtn.addEventListener('click', () => {
                scale = 1;
                translateX = 0;
                translateY = 0;
                updateTransform();
                });
            }

            if (fullscreenBtn) {
                fullscreenBtn.addEventListener('click', () => {
                if (container.requestFullscreen) {
                    container.requestFullscreen();
                } else if (container.webkitRequestFullscreen) {
                    container.webkitRequestFullscreen();
                } else if (container.msRequestFullscreen) {
                    container.msRequestFullscreen();
                }
                });
            }

            // Mouse Events
            mermaidElement.addEventListener('mousedown', (e) => {
                if (isTouch) return; // 如果是触摸设备，忽略鼠标事件

                isDragging = true;
                startX = e.clientX - translateX;
                startY = e.clientY - translateY;
                mermaidElement.style.cursor = 'grabbing';
                updateTransform();
                e.preventDefault();
            });

            document.addEventListener('mousemove', (e) => {
                if (isDragging && !isTouch) {
                translateX = e.clientX - startX;
                translateY = e.clientY - startY;
                updateTransform();
                }
            });

            document.addEventListener('mouseup', () => {
                if (isDragging && !isTouch) {
                isDragging = false;
                mermaidElement.style.cursor = 'grab';
                updateTransform();
                }
            });

            document.addEventListener('mouseleave', () => {
                if (isDragging && !isTouch) {
                isDragging = false;
                mermaidElement.style.cursor = 'grab';
                updateTransform();
                }
            });

            // 获取两点之间的距离
            function getTouchDistance(touch1, touch2) {
                return Math.hypot(
                touch2.clientX - touch1.clientX,
                touch2.clientY - touch1.clientY
                );
            }

            // Touch Events - 触摸事件处理
            mermaidElement.addEventListener('touchstart', (e) => {
                isTouch = true;
                touchStartTime = Date.now();

                if (e.touches.length === 1) {
                // 单指拖动
                isPinching = false;
                isDragging = true;

                const touch = e.touches[0];
                startX = touch.clientX - translateX;
                startY = touch.clientY - translateY;

                } else if (e.touches.length === 2) {
                // 双指缩放
                isPinching = true;
                isDragging = false;

                const touch1 = e.touches[0];
                const touch2 = e.touches[1];
                initialDistance = getTouchDistance(touch1, touch2);
                initialScale = scale;
                }

                e.preventDefault();
            }, { passive: false });

            mermaidElement.addEventListener('touchmove', (e) => {
                if (e.touches.length === 1 && isDragging && !isPinching) {
                // 单指拖动
                const touch = e.touches[0];
                translateX = touch.clientX - startX;
                translateY = touch.clientY - startY;
                updateTransform();

                } else if (e.touches.length === 2 && isPinching) {
                // 双指缩放
                const touch1 = e.touches[0];
                const touch2 = e.touches[1];
                const currentDistance = getTouchDistance(touch1, touch2);

                if (initialDistance > 0) {
                    const newScale = Math.min(Math.max(
                    initialScale * (currentDistance / initialDistance),
                    0.3
                    ), 4);
                    scale = newScale;
                    updateTransform();
                }
                }

                e.preventDefault();
            }, { passive: false });

            mermaidElement.addEventListener('touchend', (e) => {
                // 重置状态
                if (e.touches.length === 0) {
                isDragging = false;
                isPinching = false;
                initialDistance = 0;

                // 延迟重置isTouch，避免鼠标事件立即触发
                setTimeout(() => {
                    isTouch = false;
                }, 100);
                } else if (e.touches.length === 1 && isPinching) {
                // 从双指变为单指，切换为拖动模式
                isPinching = false;
                isDragging = true;

                const touch = e.touches[0];
                startX = touch.clientX - translateX;
                startY = touch.clientY - translateY;
                }

                updateTransform();
            });

            mermaidElement.addEventListener('touchcancel', (e) => {
                isDragging = false;
                isPinching = false;
                initialDistance = 0;

                setTimeout(() => {
                isTouch = false;
                }, 100);

                updateTransform();
            });

            // Enhanced wheel zoom with better center point handling
            container.addEventListener('wheel', (e) => {
                e.preventDefault();
                const rect = container.getBoundingClientRect();
                const centerX = rect.width / 2;
                const centerY = rect.height / 2;

                const delta = e.deltaY > 0 ? 0.9 : 1.1;
                const newScale = Math.min(Math.max(scale * delta, 0.3), 4);

                // Adjust translation to zoom towards center
                if (newScale !== scale) {
                const scaleDiff = newScale / scale;
                translateX = translateX * scaleDiff;
                translateY = translateY * scaleDiff;
                scale = newScale;

                if (scale <= 1) {
                    translateX = 0;
                    translateY = 0;
                }

                updateTransform();
                }
            });

            // Initialize display
            updateTransform();
            });
        }

        // Initialize mermaid controls after DOM is loaded
        document.addEventListener('DOMContentLoaded', function() {
            initializeMermaidControls();
        });

        // Smooth scrolling for anchor links
        document.querySelectorAll('a[href^="#"]').forEach(anchor => {
            anchor.addEventListener('click', function (e) {
                e.preventDefault();
                const target = document.querySelector(this.getAttribute('href'));
                if (target) {
                    target.scrollIntoView({
                        behavior: 'smooth',
                        block: 'start'
                    });
                }
            });
        });

        // Active TOC highlighting
        function updateActiveToc() {
            const sections = document.querySelectorAll('section[id]');
            const tocItems = document.querySelectorAll('.toc-item');
            
            let currentSection = '';
            sections.forEach(section => {
                const rect = section.getBoundingClientRect();
                if (rect.top <= 100 && rect.bottom >= 100) {
                    currentSection = section.id;
                }
            });
            
            tocItems.forEach(item => {
                item.classList.remove('active');
                if (item.getAttribute('href') === '#' + currentSection) {
                    item.classList.add('active');
                }
            });
        }

        window.addEventListener('scroll', updateActiveToc);
        updateActiveToc(); // Call once on load

        // Toggle TOC on mobile
        const tocToggle = document.getElementById('tocToggle');
        const tocNav = document.getElementById('tocNav');
        let isTocActive = false;

        tocToggle.addEventListener('click', function() {
            isTocActive = !isTocActive;
            tocNav.classList.toggle('active');
            document.body.classList.toggle('toc-active');
            
            // Change icon
            const icon = this.querySelector('i');
            icon.classList.toggle('fa-bars');
            icon.classList.toggle('fa-times');
        });

        // Close TOC when clicking on a link (mobile)
        document.querySelectorAll('.toc-item').forEach(link => {
            link.addEventListener('click', function() {
                if (window.innerWidth <= 1024) {
                    isTocActive = false;
                    tocNav.classList.remove('active');
                    document.body.classList.remove('toc-active');
                    
                    // Reset icon
                    const icon = tocToggle.querySelector('i');
                    icon.classList.add('fa-bars');
                    icon.classList.remove('fa-times');
                }
            });
        });

        // Handle window resize to remove active TOC on large screens
        function handleResize() {
            if (window.innerWidth > 1024 && isTocActive) {
                isTocActive = false;
                tocNav.classList.remove('active');
                document.body.classList.remove('toc-active');
                
                const icon = tocToggle.querySelector('i');
                icon.classList.add('fa-bars');
                icon.classList.remove('fa-times');
            }
        }

        window.addEventListener('resize', handleResize);
    </script>
  

</body></html>                    

讨论回复

1 条回复

✨步子哥 (steper) #1

02-21 12:42

                                        <html><body>
<!-- Hero Section -->
      <section class="bg-gradient-to-br from-slate-50 to-blue-50 py-16 px-8">
        <div class="max-w-6xl mx-auto">
          <div class="hero-grid">
            <!-- Hero Text -->
            <div class="space-y-6">
              <div class="text-sm font-medium text-blue-600 tracking-wide uppercase">
                技术深度分析报告
              </div>
              <h1 class="serif text-5xl lg:text-6xl font-bold leading-tight hero-text">
                GLM-5：
                <br/>
                <em class="text-4xl lg:text-5xl">开源Agentic Engineering的新纪元</em>
              </h1>
              <p class="text-xl text-gray-600 leading-relaxed">
                7440亿参数的稀疏MoE架构，DeepSeek注意力机制，
                <br/>
                以及异步强化学习框架Slime的技术突破
              </p>
              <div class="flex flex-wrap gap-3 pt-4">
                <span class="px-4 py-2 bg-blue-100 text-blue-800 rounded-full text-sm font-medium">
                  <i class="fas fa-brain mr-2"></i>744B 参数
                </span>
                <span class="px-4 py-2 bg-green-100 text-green-800 rounded-full text-sm font-medium">
                  <i class="fas fa-code mr-2"></i>77.8% SWE-bench
                </span>
                <span class="px-4 py-2 bg-purple-100 text-purple-800 rounded-full text-sm font-medium">
                  <i class="fas fa-unlock mr-2"></i>MIT开源
                </span>
              </div>
            </div>

            <!-- Hero Visual -->
            <div class="relative">
              <img src="https://kimi-web-img.moonshot.cn/img/s6.51cto.com/2161ecc1780e781973d84eb42de6b0a2cae45667.webp" alt="AI神经网络架构示意图" class="w-full h-80 object-cover rounded-2xl shadow-2xl" size="medium" aspect="wide" query="AI神经网络架构" referrerpolicy="no-referrer" data-modified="1" data-score="0.00"/>
              <div class="absolute inset-0 bg-gradient-to-t from-black/30 to-transparent rounded-2xl"></div>
              <div class="absolute bottom-4 left-4 text-white">
                <div class="text-sm font-medium">稀疏MoE架构 · 异步RL训练</div>
              </div>
            </div>
          </div>
        </div>
      </section>                                    

需要登录才能发表回复

登录注册

GLM-5： 开源Agentic Engineering的新纪元

讨论回复

推荐

GLM-5：开源Agentic Engineering的新纪元