feat(blog): 添加通义千问Qwen3.5深度分析文章及配套图表
- 新增关于Qwen3.5登顶中国第一的深度技术解读文章 - 添加LMArena评测机制可视化图表页面 - 添加Qwen3.5 MoE混合专家架构示意图 - 添加综合能力跑分对比图表 - 添加价格对比可视化页面 - 添加使用场景推荐图表 - 更新.gitignore文件添加.playwright-mcp目录忽略规则
3
.gitignore
vendored
@ -1,2 +1,3 @@
|
|||||||
/.idea/
|
/.idea/
|
||||||
/项目/
|
/项目/
|
||||||
|
/.playwright-mcp/
|
||||||
|
|||||||
251
articles/007-国产模型杀疯了!通义千问Qwen3.5登顶中国第一,凭什么超越GPT-5.4.md
Normal file
@ -0,0 +1,251 @@
|
|||||||
|
# 国产模型杀疯了!通义千问 Qwen3.5 登顶中国第一,凭什么超越 GPT-5.4?
|
||||||
|
|
||||||
|
> 发布日期:2026-03-29
|
||||||
|
> 分类:技术解读 / 深度分析
|
||||||
|
> 作者:老邓唠AI
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
## 引子:中国模型,第一次站到了这个位置
|
||||||
|
|
||||||
|
3 月 20 日,阿里悄悄在全球最权威的大模型竞技平台 LMArena 上提交了一个新模型——**Qwen3.5-Max-Preview**。
|
||||||
|
|
||||||
|
结果一出,整个 AI 圈炸了。
|
||||||
|
|
||||||
|
**1464 分,全球公司排名第五,中国第一。**
|
||||||
|
|
||||||
|
它超越了 OpenAI 的 GPT-5.4、xAI 的 Grok 4.1、字节的豆包 2.0、智谱的 GLM-5、月之暗面的 Kimi 2.5——几乎打遍了国内外一众顶尖选手。
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
要知道,这还只是一个**预览版**。正式版还没发。
|
||||||
|
|
||||||
|
今天就来拆解一下:通义千问 Qwen3.5 到底强在哪?这个"中国第一"的含金量有多高?普通人又能怎么用?
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 一、LMArena 是什么?为什么它的排名最有说服力?
|
||||||
|
|
||||||
|
在聊 Qwen3.5 之前,先说说 LMArena 这个平台——因为不是所有排行榜都值得看。
|
||||||
|
|
||||||
|
### 传统跑分的问题
|
||||||
|
|
||||||
|
我们常见的 MMLU、HumanEval、GPQA 这些基准测试,本质上是**标准化考试**。模型厂商可以针对性训练、刷分、甚至"背题"。就像学生刷模拟卷能拿高分,但不代表真实水平。
|
||||||
|
|
||||||
|
### LMArena 的不同之处
|
||||||
|
|
||||||
|
LMArena(原 LMSYS Chatbot Arena)采用的是一套完全不同的评测机制:
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
1. **匿名对战**:两个模型的回答并排展示,用户不知道哪个是谁
|
||||||
|
2. **真人盲测**:全球开发者根据回答质量投票选出胜者
|
||||||
|
3. **ELO 评分**:采用国际象棋的 ELO 积分系统,胜率越高分数越高
|
||||||
|
4. **规模庞大**:截至 2026 年 3 月,已有超过 **563 万次投票**,覆盖 **333 个模型**
|
||||||
|
|
||||||
|
简单说,LMArena 不是模型自己考试,而是**让全球用户当裁判,真刀真枪地 PK**。这就是为什么它被公认为最权威的大模型评测平台。
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 二、当前全球大模型排行格局
|
||||||
|
|
||||||
|
先看大盘。截至 2026 年 3 月底,LMArena 文本排行榜的竞争格局如下:
|
||||||
|
|
||||||
|
### LMArena 全球 Top 模型排名
|
||||||
|
|
||||||
|
| 排名 | 模型 | 公司 | ELO 分数 | 亮点 |
|
||||||
|
|------|------|------|---------|------|
|
||||||
|
| 1 | Claude Opus 4.6 | Anthropic | 1504 | 综合最强,编程之王 |
|
||||||
|
| 2 | Gemini 3.1 Pro Preview | Google | 1500 | 科学推理 GPQA 94.3% 史上最高 |
|
||||||
|
| 3 | Claude Opus 4.6 Thinking | Anthropic | ~1500 | 推理增强版 |
|
||||||
|
| 4 | Grok 4.20 Beta | xAI | 1493 | 马斯克旗下,进步飞快 |
|
||||||
|
| 5 | Gemini 3 Pro | Google | 1485 | 多模态标杆 |
|
||||||
|
| 6 | GPT-5.4 Thinking | OpenAI | — | Agent 能力超越人类基线 |
|
||||||
|
| ... | ... | ... | ... | ... |
|
||||||
|
| **中国第一** | **Qwen3.5-Max-Preview** | **阿里** | **1464** | **全球公司排名第五** |
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
### 关键结论:没有绝对的"最强"
|
||||||
|
|
||||||
|
现在的大模型竞争已经进入**多极化时代**,各家都有自己的长板:
|
||||||
|
|
||||||
|
- **Claude Opus 4.6**:LMArena 综合第一,SWE-Bench 编程 80.9%,代码工程最强
|
||||||
|
- **Gemini 3.1 Pro**:科学推理 GPQA 94.3%,学术研究首选
|
||||||
|
- **GPT-5.4**:OSWorld 75% 桌面操作超越人类,Agent 能力最强
|
||||||
|
- **Qwen3.5**:开源最强 + 性价比之王 + 原生多模态,中国第一
|
||||||
|
|
||||||
|
而 Qwen3.5 的位置非常特殊——它是这个顶级梯队里**唯一的开源模型**,也是**唯一的中国模型**。
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 三、Qwen3.5 到底强在哪?
|
||||||
|
|
||||||
|
现在来拆解 Qwen3.5 本身。它能拿到这个排名,不是靠运气,而是有实打实的技术创新。
|
||||||
|
|
||||||
|
### 3.1 全球首个原生多模态 MoE 大模型
|
||||||
|
|
||||||
|
Qwen3.5-397B-A17B 这个型号名已经说明了一切:
|
||||||
|
|
||||||
|
- **397B**:总参数 **3970 亿**
|
||||||
|
- **A17B**:每次推理仅激活 **170 亿** 参数
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
这就是 **MoE(混合专家)** 架构的威力——模型很大,但每次只调用最相关的"专家"来回答问题,**用不到 5% 的计算量实现接近全量模型的效果**。
|
||||||
|
|
||||||
|
更关键的是,Qwen3.5 在 MoE 的基础上做了两个创新:
|
||||||
|
|
||||||
|
1. **线性注意力(Gated Delta Networks)**:大幅降低长文本处理的计算开销
|
||||||
|
2. **原生多模态预训练**:从一开始就用文本+图像+视频混合训练,不是后期拼接
|
||||||
|
|
||||||
|
结果就是:**不到 40% 的参数量,达到了上一代万亿参数 Qwen3-Max 的性能水平**。
|
||||||
|
|
||||||
|
### 3.2 跑分数据全面开花
|
||||||
|
|
||||||
|
来看硬核数据:
|
||||||
|
|
||||||
|
#### 综合能力对比
|
||||||
|
|
||||||
|
| 基准测试 | Qwen3.5 | GPT-5.2 | Claude 4.5 | 说明 |
|
||||||
|
|---------|---------|---------|-----------|------|
|
||||||
|
| MMLU-Pro | **87.8** | 86.5 | — | 知识推理,超越 GPT-5.2 |
|
||||||
|
| GPQA | **88.4** | — | 87.9 | 博士级难题,超越 Claude 4.5 |
|
||||||
|
| AIME 2026 | 91.3 | **96.7** | 93.3 | 数学竞赛,仍有差距 |
|
||||||
|
| IFBench | **76.5** | 75.4 | 58.0 | 指令遵循,大幅领先 |
|
||||||
|
|
||||||
|
#### LMArena 细分排名
|
||||||
|
|
||||||
|
| 维度 | 全球排名 | 中国排名 |
|
||||||
|
|------|---------|---------|
|
||||||
|
| 综合(无风格控制) | 第 6 | 第 1 |
|
||||||
|
| 数学能力 | 第 5 | 第 1 |
|
||||||
|
| 专家文本 | 第 10 | 第 1 |
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
### 3.3 原生多模态:不只是看图,还能看视频
|
||||||
|
|
||||||
|
Qwen3.5 的多模态能力不是"加个视觉模块"那么简单。它从预训练阶段就把视觉和文本放在一起训练,属于**早期融合(Early Fusion)**。
|
||||||
|
|
||||||
|
能做什么:
|
||||||
|
|
||||||
|
- **图像理解**:读文档、看图表、识别 UI 界面
|
||||||
|
- **视频分析**:支持最长 **2 小时** 的视频直接输入
|
||||||
|
- **GUI 智能体**:自主操作手机和电脑界面完成任务
|
||||||
|
- **支持 201 种语言**
|
||||||
|
|
||||||
|
在多模态推理、视觉问答、文本识别、空间智能、视频理解等评测中,Qwen3.5 均拿到了开源模型的最佳成绩。
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 四、价格屠夫:百万 Token 只要 0.8 元
|
||||||
|
|
||||||
|
性能强是一方面,但 Qwen3.5 真正让人兴奋的,是它的**价格和开源策略**。
|
||||||
|
|
||||||
|
### API 定价对比
|
||||||
|
|
||||||
|
| 模型 | 百万输入 Token 价格 | 百万输出 Token 价格 |
|
||||||
|
|------|-------------------|-------------------|
|
||||||
|
| GPT-5.4 | ≈ ¥15 | ≈ ¥60 |
|
||||||
|
| Claude Opus 4.6 | ≈ ¥15 | ≈ ¥75 |
|
||||||
|
| Gemini 3 Pro | ≈ ¥14.4 | ≈ ¥57.6 |
|
||||||
|
| **Qwen3.5-Plus** | **¥0.8** | **¥— (极低)** |
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
没看错,**Qwen3.5-Plus 的输入价格是 Gemini 3 Pro 的 1/18**。
|
||||||
|
|
||||||
|
### 部署效率
|
||||||
|
|
||||||
|
和上一代 Qwen3-Max 相比:
|
||||||
|
|
||||||
|
- 显存占用**降低 60%**
|
||||||
|
- 推理吞吐量**最高提升 19 倍**
|
||||||
|
|
||||||
|
这意味着同样的 GPU 资源,能服务更多用户,成本进一步摊薄。
|
||||||
|
|
||||||
|
### 开源优势
|
||||||
|
|
||||||
|
Qwen3.5 全系列开源,这意味着:
|
||||||
|
|
||||||
|
- **本地部署**:可以在自己的服务器上跑,数据不出境
|
||||||
|
- **可微调**:针对特定业务场景定制模型
|
||||||
|
- **无 API 依赖**:不用担心服务商涨价或停服
|
||||||
|
- **社区生态**:Ollama、vLLM、ModelScope 等工具链全面支持
|
||||||
|
|
||||||
|
对于企业用户来说,这可能比排行榜上的分数更有吸引力。
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 五、普通人怎么用?
|
||||||
|
|
||||||
|
说了这么多技术细节,落到实际使用上,有几个入口可以体验 Qwen3.5:
|
||||||
|
|
||||||
|
### 5.1 最简单:通义千问官网
|
||||||
|
|
||||||
|
直接访问 [tongyi.aliyun.com](https://tongyi.aliyun.com) 或者下载通义千问 App,就能免费对话体验。
|
||||||
|
|
||||||
|
### 5.2 开发者:阿里云百炼 API
|
||||||
|
|
||||||
|
通过阿里云百炼平台接入 API,适合需要集成到自己产品里的开发者。定价极具竞争力,还有新用户免费额度。
|
||||||
|
|
||||||
|
### 5.3 本地部署:中等规模模型
|
||||||
|
|
||||||
|
如果你有一张不错的显卡,可以用 Ollama 或 vLLM 部署 Qwen3.5 的中等规模模型(4B、9B、27B):
|
||||||
|
|
||||||
|
```bash
|
||||||
|
# 以 Ollama 为例
|
||||||
|
ollama run qwen3.5:27b
|
||||||
|
```
|
||||||
|
|
||||||
|
2B 和 4B 模型甚至可以在手机端运行,适合端侧部署。
|
||||||
|
|
||||||
|
### 5.4 适合什么场景?
|
||||||
|
|
||||||
|
| 场景 | 推荐模型 | 说明 |
|
||||||
|
|------|---------|------|
|
||||||
|
| 日常对话、写作 | Qwen3.5-Plus | 性价比最高 |
|
||||||
|
| 编程辅助 | Qwen3.5-Max | 代码能力强 |
|
||||||
|
| 图片/文档分析 | Qwen3.5(多模态) | 原生视觉理解 |
|
||||||
|
| 视频内容理解 | Qwen3.5(多模态) | 支持 2 小时视频 |
|
||||||
|
| 企业私有化部署 | Qwen3.5-27B 开源版 | 数据不出境 |
|
||||||
|
| 手机端 AI 助手 | Qwen3.5-2B | 轻量本地运行 |
|
||||||
|
|
||||||
|

|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
## 写在最后:里程碑,但保持清醒
|
||||||
|
|
||||||
|
Qwen3.5-Max-Preview 在 LMArena 上的表现,确实是中国大模型发展的一个**里程碑时刻**。
|
||||||
|
|
||||||
|
从两年前的追赶,到今天能在全球最权威的盲测平台上超越 GPT-5.4、比肩 Gemini 和 Claude——这个进步是实实在在的。
|
||||||
|
|
||||||
|
但也要保持清醒:
|
||||||
|
|
||||||
|
1. **这是预览版**,正式版的表现还需要观察
|
||||||
|
2. **LMArena 测的是综合对话体验**,在编程(SWE-Bench)、数学(AIME)等单项上,Qwen3.5 和顶尖模型还有差距
|
||||||
|
3. **排行榜分数 ≠ 实际体验**,真正好不好用,还得自己试
|
||||||
|
|
||||||
|
但不管怎么说,**中国终于有了一个能在全球顶尖梯队里站稳脚跟的大模型**。而且它还是开源的、便宜的、可以本地部署的。
|
||||||
|
|
||||||
|
这对整个中国 AI 生态来说,意义重大。
|
||||||
|
|
||||||
|
期待 Qwen3.5 正式版的表现。
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
> **如果这篇文章对你有帮助,欢迎点赞、收藏、转发,你的支持是我持续输出的最大动力。**
|
||||||
|
>
|
||||||
|
> 关注「老邓唠AI」,每周带你拆解 AI 圈最值得关注的大事。
|
||||||
|
|
||||||
|
---
|
||||||
|
|
||||||
|
**参考来源:**
|
||||||
|
|
||||||
|
- [LMArena 官方排行榜](https://arena.ai/leaderboard/text)
|
||||||
|
- [阿里云通义千问官网](https://www.aliyun.com/product/tongyi)
|
||||||
|
- [Qwen3.5 技术博客](https://developer.aliyun.com/article/1712860)
|
||||||
|
- [IT之家:阿里通义千问 3.5-Max-Preview 首发亮相](https://www.ithome.com/0/930/834.htm)
|
||||||
|
- [量子位:Qwen3.5-Max 预览版首度亮相](https://www.qbitai.com/2026/03/389610.html)
|
||||||
BIN
articles/007/benchmark-comparison.png
Normal file
|
After Width: | Height: | Size: 266 KiB |
BIN
articles/007/cover.png
Normal file
|
After Width: | Height: | Size: 820 KiB |
BIN
articles/007/cover_raw.png
Normal file
|
After Width: | Height: | Size: 1.2 MiB |
69
articles/007/diagrams/benchmark-comparison.html
Normal file
@ -0,0 +1,69 @@
|
|||||||
|
<!DOCTYPE html>
|
||||||
|
<html lang="zh-CN">
|
||||||
|
<head>
|
||||||
|
<meta charset="UTF-8">
|
||||||
|
<style>
|
||||||
|
* { margin: 0; padding: 0; box-sizing: border-box; }
|
||||||
|
body { width: 1200px; height: 680px; background: linear-gradient(135deg, #0f0c29, #302b63, #24243e); font-family: -apple-system, "PingFang SC", "Microsoft YaHei", sans-serif; display: flex; align-items: center; justify-content: center; color: #fff; }
|
||||||
|
.container { width: 1080px; }
|
||||||
|
h2 { text-align: center; font-size: 32px; margin-bottom: 36px; background: linear-gradient(90deg, #a78bfa, #60a5fa); -webkit-background-clip: text; -webkit-text-fill-color: transparent; }
|
||||||
|
.chart-group { display: flex; flex-direction: column; gap: 24px; }
|
||||||
|
.chart-row { display: flex; align-items: center; gap: 16px; }
|
||||||
|
.chart-label { width: 130px; text-align: right; font-size: 15px; font-weight: 600; color: rgba(255,255,255,0.8); flex-shrink: 0; }
|
||||||
|
.bars { flex: 1; display: flex; flex-direction: column; gap: 6px; }
|
||||||
|
.bar-item { display: flex; align-items: center; gap: 10px; }
|
||||||
|
.bar-name { width: 100px; font-size: 12px; color: rgba(255,255,255,0.6); text-align: right; flex-shrink: 0; }
|
||||||
|
.bar-track { flex: 1; height: 28px; background: rgba(255,255,255,0.05); border-radius: 6px; position: relative; overflow: hidden; }
|
||||||
|
.bar-fill { height: 100%; border-radius: 6px; display: flex; align-items: center; justify-content: flex-end; padding-right: 10px; font-size: 13px; font-weight: 700; }
|
||||||
|
.bar-fill.qwen { background: linear-gradient(90deg, #7c3aed, #a78bfa); }
|
||||||
|
.bar-fill.gpt { background: linear-gradient(90deg, #047857, #34d399); }
|
||||||
|
.bar-fill.claude { background: linear-gradient(90deg, #b45309, #fbbf24); }
|
||||||
|
.bar-fill.best { box-shadow: 0 0 12px rgba(167,139,250,0.5); }
|
||||||
|
.legend { display: flex; gap: 24px; justify-content: center; margin-bottom: 20px; }
|
||||||
|
.legend-item { display: flex; align-items: center; gap: 6px; font-size: 13px; color: rgba(255,255,255,0.7); }
|
||||||
|
.legend-dot { width: 12px; height: 12px; border-radius: 3px; }
|
||||||
|
</style>
|
||||||
|
</head>
|
||||||
|
<body>
|
||||||
|
<div class="container">
|
||||||
|
<h2>Qwen3.5 综合能力跑分对比</h2>
|
||||||
|
<div class="legend">
|
||||||
|
<div class="legend-item"><div class="legend-dot" style="background:#a78bfa;"></div>Qwen3.5</div>
|
||||||
|
<div class="legend-item"><div class="legend-dot" style="background:#34d399;"></div>GPT-5.2</div>
|
||||||
|
<div class="legend-item"><div class="legend-dot" style="background:#fbbf24;"></div>Claude 4.5</div>
|
||||||
|
</div>
|
||||||
|
<div class="chart-group">
|
||||||
|
<div class="chart-row">
|
||||||
|
<div class="chart-label">MMLU-Pro<br><span style="font-size:11px;color:rgba(255,255,255,0.4);">知识推理</span></div>
|
||||||
|
<div class="bars">
|
||||||
|
<div class="bar-item"><div class="bar-name">Qwen3.5</div><div class="bar-track"><div class="bar-fill qwen best" style="width:92%;">87.8</div></div></div>
|
||||||
|
<div class="bar-item"><div class="bar-name">GPT-5.2</div><div class="bar-track"><div class="bar-fill gpt" style="width:90.5%;">86.5</div></div></div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
<div class="chart-row">
|
||||||
|
<div class="chart-label">GPQA<br><span style="font-size:11px;color:rgba(255,255,255,0.4);">博士级难题</span></div>
|
||||||
|
<div class="bars">
|
||||||
|
<div class="bar-item"><div class="bar-name">Qwen3.5</div><div class="bar-track"><div class="bar-fill qwen best" style="width:91%;">88.4</div></div></div>
|
||||||
|
<div class="bar-item"><div class="bar-name">Claude 4.5</div><div class="bar-track"><div class="bar-fill claude" style="width:90.5%;">87.9</div></div></div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
<div class="chart-row">
|
||||||
|
<div class="chart-label">AIME 2026<br><span style="font-size:11px;color:rgba(255,255,255,0.4);">数学竞赛</span></div>
|
||||||
|
<div class="bars">
|
||||||
|
<div class="bar-item"><div class="bar-name">Qwen3.5</div><div class="bar-track"><div class="bar-fill qwen" style="width:91.3%;">91.3</div></div></div>
|
||||||
|
<div class="bar-item"><div class="bar-name">GPT-5.2</div><div class="bar-track"><div class="bar-fill gpt best" style="width:96.7%;">96.7</div></div></div>
|
||||||
|
<div class="bar-item"><div class="bar-name">Claude 4.5</div><div class="bar-track"><div class="bar-fill claude" style="width:93.3%;">93.3</div></div></div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
<div class="chart-row">
|
||||||
|
<div class="chart-label">IFBench<br><span style="font-size:11px;color:rgba(255,255,255,0.4);">指令遵循</span></div>
|
||||||
|
<div class="bars">
|
||||||
|
<div class="bar-item"><div class="bar-name">Qwen3.5</div><div class="bar-track"><div class="bar-fill qwen best" style="width:85%;">76.5</div></div></div>
|
||||||
|
<div class="bar-item"><div class="bar-name">GPT-5.2</div><div class="bar-track"><div class="bar-fill gpt" style="width:83.8%;">75.4</div></div></div>
|
||||||
|
<div class="bar-item"><div class="bar-name">Claude 4.5</div><div class="bar-track"><div class="bar-fill claude" style="width:64.4%;">58.0</div></div></div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
</body>
|
||||||
|
</html>
|
||||||
51
articles/007/diagrams/lmarena-mechanism.html
Normal file
@ -0,0 +1,51 @@
|
|||||||
|
<!DOCTYPE html>
|
||||||
|
<html lang="zh-CN">
|
||||||
|
<head>
|
||||||
|
<meta charset="UTF-8">
|
||||||
|
<style>
|
||||||
|
* { margin: 0; padding: 0; box-sizing: border-box; }
|
||||||
|
body { width: 1200px; height: 680px; background: linear-gradient(135deg, #0f0c29, #302b63, #24243e); font-family: -apple-system, "PingFang SC", "Microsoft YaHei", sans-serif; display: flex; align-items: center; justify-content: center; color: #fff; }
|
||||||
|
.container { width: 1100px; }
|
||||||
|
h2 { text-align: center; font-size: 32px; margin-bottom: 48px; background: linear-gradient(90deg, #a78bfa, #60a5fa); -webkit-background-clip: text; -webkit-text-fill-color: transparent; }
|
||||||
|
.flow { display: flex; align-items: center; justify-content: center; gap: 16px; }
|
||||||
|
.step { background: rgba(255,255,255,0.08); border: 1px solid rgba(255,255,255,0.15); border-radius: 16px; padding: 28px 24px; width: 220px; text-align: center; backdrop-filter: blur(10px); }
|
||||||
|
.step .icon { font-size: 48px; margin-bottom: 12px; }
|
||||||
|
.step .title { font-size: 18px; font-weight: 700; margin-bottom: 8px; color: #c4b5fd; }
|
||||||
|
.step .desc { font-size: 13px; line-height: 1.6; color: rgba(255,255,255,0.7); }
|
||||||
|
.arrow { font-size: 32px; color: #60a5fa; flex-shrink: 0; }
|
||||||
|
.bottom-note { text-align: center; margin-top: 40px; font-size: 15px; color: rgba(255,255,255,0.5); }
|
||||||
|
.bottom-note span { color: #fbbf24; font-weight: 700; font-size: 20px; }
|
||||||
|
</style>
|
||||||
|
</head>
|
||||||
|
<body>
|
||||||
|
<div class="container">
|
||||||
|
<h2>LMArena 评测机制:全球用户当裁判</h2>
|
||||||
|
<div class="flow">
|
||||||
|
<div class="step">
|
||||||
|
<div class="icon">👤</div>
|
||||||
|
<div class="title">用户提问</div>
|
||||||
|
<div class="desc">用户输入一个真实问题,系统随机分配两个模型</div>
|
||||||
|
</div>
|
||||||
|
<div class="arrow">→</div>
|
||||||
|
<div class="step">
|
||||||
|
<div class="icon">🤖🤖</div>
|
||||||
|
<div class="title">匿名对战</div>
|
||||||
|
<div class="desc">两个模型同时回答,用户不知道哪个是谁</div>
|
||||||
|
</div>
|
||||||
|
<div class="arrow">→</div>
|
||||||
|
<div class="step">
|
||||||
|
<div class="icon">👆</div>
|
||||||
|
<div class="title">真人盲测</div>
|
||||||
|
<div class="desc">用户根据回答质量投票选出更好的那个</div>
|
||||||
|
</div>
|
||||||
|
<div class="arrow">→</div>
|
||||||
|
<div class="step">
|
||||||
|
<div class="icon">📊</div>
|
||||||
|
<div class="title">ELO 评分</div>
|
||||||
|
<div class="desc">采用国际象棋 ELO 积分系统,胜率越高分数越高</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
<div class="bottom-note">截至 2026 年 3 月,已累计 <span>563 万+</span> 次投票,覆盖 <span>333</span> 个模型</div>
|
||||||
|
</div>
|
||||||
|
</body>
|
||||||
|
</html>
|
||||||
70
articles/007/diagrams/moe-architecture.html
Normal file
@ -0,0 +1,70 @@
|
|||||||
|
<!DOCTYPE html>
|
||||||
|
<html lang="zh-CN">
|
||||||
|
<head>
|
||||||
|
<meta charset="UTF-8">
|
||||||
|
<style>
|
||||||
|
* { margin: 0; padding: 0; box-sizing: border-box; }
|
||||||
|
body { width: 1200px; height: 680px; background: linear-gradient(135deg, #0f0c29, #302b63, #24243e); font-family: -apple-system, "PingFang SC", "Microsoft YaHei", sans-serif; display: flex; align-items: center; justify-content: center; color: #fff; }
|
||||||
|
.container { width: 1100px; }
|
||||||
|
h2 { text-align: center; font-size: 32px; margin-bottom: 40px; background: linear-gradient(90deg, #a78bfa, #60a5fa); -webkit-background-clip: text; -webkit-text-fill-color: transparent; }
|
||||||
|
.main { display: flex; gap: 40px; align-items: center; justify-content: center; }
|
||||||
|
.left { display: flex; flex-direction: column; gap: 16px; align-items: center; }
|
||||||
|
.param-box { background: rgba(167,139,250,0.15); border: 2px solid #a78bfa; border-radius: 16px; padding: 20px 32px; text-align: center; }
|
||||||
|
.param-box .big { font-size: 48px; font-weight: 800; color: #a78bfa; }
|
||||||
|
.param-box .label { font-size: 14px; color: rgba(255,255,255,0.6); margin-top: 4px; }
|
||||||
|
.center-visual { position: relative; width: 360px; height: 360px; }
|
||||||
|
.expert { position: absolute; width: 80px; height: 80px; border-radius: 50%; display: flex; align-items: center; justify-content: center; font-size: 13px; font-weight: 600; border: 2px solid rgba(255,255,255,0.2); background: rgba(255,255,255,0.05); color: rgba(255,255,255,0.4); transition: all 0.3s; }
|
||||||
|
.expert.active { border-color: #34d399; background: rgba(52,211,153,0.2); color: #34d399; box-shadow: 0 0 20px rgba(52,211,153,0.3); }
|
||||||
|
.router { position: absolute; top: 50%; left: 50%; transform: translate(-50%, -50%); width: 100px; height: 100px; border-radius: 50%; background: linear-gradient(135deg, #6366f1, #8b5cf6); display: flex; align-items: center; justify-content: center; font-size: 14px; font-weight: 700; box-shadow: 0 0 30px rgba(99,102,241,0.4); }
|
||||||
|
.right { display: flex; flex-direction: column; gap: 16px; }
|
||||||
|
.feat { background: rgba(255,255,255,0.06); border: 1px solid rgba(255,255,255,0.1); border-radius: 12px; padding: 16px 20px; width: 280px; }
|
||||||
|
.feat .ft { font-size: 15px; font-weight: 700; color: #60a5fa; margin-bottom: 4px; }
|
||||||
|
.feat .fd { font-size: 13px; color: rgba(255,255,255,0.6); line-height: 1.5; }
|
||||||
|
.bottom { text-align: center; margin-top: 32px; font-size: 16px; color: #fbbf24; font-weight: 600; }
|
||||||
|
</style>
|
||||||
|
</head>
|
||||||
|
<body>
|
||||||
|
<div class="container">
|
||||||
|
<h2>Qwen3.5 MoE 混合专家架构</h2>
|
||||||
|
<div class="main">
|
||||||
|
<div class="left">
|
||||||
|
<div class="param-box">
|
||||||
|
<div class="big">3970 亿</div>
|
||||||
|
<div class="label">总参数量 (397B)</div>
|
||||||
|
</div>
|
||||||
|
<div style="font-size: 28px; color: #60a5fa;">↓ 仅激活</div>
|
||||||
|
<div class="param-box" style="border-color: #34d399; background: rgba(52,211,153,0.15);">
|
||||||
|
<div class="big" style="color: #34d399;">170 亿</div>
|
||||||
|
<div class="label">激活参数 (A17B) ≈ 4.3%</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
<div class="center-visual">
|
||||||
|
<div class="router">路由器<br>Router</div>
|
||||||
|
<div class="expert active" style="top:10px;left:140px;">专家 1</div>
|
||||||
|
<div class="expert" style="top:50px;left:260px;">专家 2</div>
|
||||||
|
<div class="expert active" style="top:150px;left:280px;">专家 3</div>
|
||||||
|
<div class="expert" style="top:250px;left:260px;">专家 4</div>
|
||||||
|
<div class="expert" style="top:280px;left:140px;">专家 5</div>
|
||||||
|
<div class="expert" style="top:250px;left:20px;">专家 6</div>
|
||||||
|
<div class="expert active" style="top:150px;left:0px;">专家 7</div>
|
||||||
|
<div class="expert" style="top:50px;left:20px;">专家 8</div>
|
||||||
|
</div>
|
||||||
|
<div class="right">
|
||||||
|
<div class="feat">
|
||||||
|
<div class="ft">线性注意力</div>
|
||||||
|
<div class="fd">Gated Delta Networks,大幅降低长文本处理的计算开销</div>
|
||||||
|
</div>
|
||||||
|
<div class="feat">
|
||||||
|
<div class="ft">原生多模态预训练</div>
|
||||||
|
<div class="fd">文本 + 图像 + 视频混合训练,非后期拼接</div>
|
||||||
|
</div>
|
||||||
|
<div class="feat">
|
||||||
|
<div class="ft">推理效率提升</div>
|
||||||
|
<div class="fd">显存降低 60%,吞吐量提升最高 19 倍</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
<div class="bottom">不到 40% 的参数量,达到上一代万亿参数 Qwen3-Max 的性能水平</div>
|
||||||
|
</div>
|
||||||
|
</body>
|
||||||
|
</html>
|
||||||
64
articles/007/diagrams/pricing-comparison.html
Normal file
@ -0,0 +1,64 @@
|
|||||||
|
<!DOCTYPE html>
|
||||||
|
<html lang="zh-CN">
|
||||||
|
<head>
|
||||||
|
<meta charset="UTF-8">
|
||||||
|
<style>
|
||||||
|
* { margin: 0; padding: 0; box-sizing: border-box; }
|
||||||
|
body { width: 1200px; height: 680px; background: linear-gradient(135deg, #0f0c29, #302b63, #24243e); font-family: -apple-system, "PingFang SC", "Microsoft YaHei", sans-serif; display: flex; align-items: center; justify-content: center; color: #fff; }
|
||||||
|
.container { width: 1000px; }
|
||||||
|
h2 { text-align: center; font-size: 32px; margin-bottom: 48px; background: linear-gradient(90deg, #a78bfa, #60a5fa); -webkit-background-clip: text; -webkit-text-fill-color: transparent; }
|
||||||
|
.price-grid { display: flex; gap: 24px; justify-content: center; }
|
||||||
|
.price-card { background: rgba(255,255,255,0.06); border: 1px solid rgba(255,255,255,0.12); border-radius: 20px; padding: 32px 28px; width: 220px; text-align: center; position: relative; }
|
||||||
|
.price-card.highlight { border-color: #a78bfa; background: rgba(167,139,250,0.1); box-shadow: 0 0 30px rgba(167,139,250,0.15); }
|
||||||
|
.price-card.highlight::before { content: "性价比之王"; position: absolute; top: -14px; left: 50%; transform: translateX(-50%); background: linear-gradient(90deg, #7c3aed, #a78bfa); padding: 4px 16px; border-radius: 20px; font-size: 12px; font-weight: 700; white-space: nowrap; }
|
||||||
|
.model-name { font-size: 18px; font-weight: 700; margin-bottom: 8px; }
|
||||||
|
.company { font-size: 13px; color: rgba(255,255,255,0.4); margin-bottom: 24px; }
|
||||||
|
.price-label { font-size: 12px; color: rgba(255,255,255,0.5); margin-bottom: 6px; }
|
||||||
|
.price-value { font-size: 36px; font-weight: 800; margin-bottom: 4px; }
|
||||||
|
.price-unit { font-size: 12px; color: rgba(255,255,255,0.4); margin-bottom: 20px; }
|
||||||
|
.price-card:nth-child(1) .price-value { color: #34d399; }
|
||||||
|
.price-card:nth-child(2) .price-value { color: #fbbf24; }
|
||||||
|
.price-card:nth-child(3) .price-value { color: #60a5fa; }
|
||||||
|
.price-card:nth-child(4) .price-value { color: #a78bfa; }
|
||||||
|
.vs { margin-top: 24px; font-size: 14px; color: rgba(255,255,255,0.5); line-height: 1.6; }
|
||||||
|
.vs strong { color: #fbbf24; }
|
||||||
|
.bottom { text-align: center; margin-top: 40px; font-size: 18px; color: #fbbf24; font-weight: 600; }
|
||||||
|
</style>
|
||||||
|
</head>
|
||||||
|
<body>
|
||||||
|
<div class="container">
|
||||||
|
<h2>百万 Token 输入价格对比(人民币)</h2>
|
||||||
|
<div class="price-grid">
|
||||||
|
<div class="price-card">
|
||||||
|
<div class="model-name">GPT-5.4</div>
|
||||||
|
<div class="company">OpenAI</div>
|
||||||
|
<div class="price-label">百万输入 Token</div>
|
||||||
|
<div class="price-value">¥15</div>
|
||||||
|
<div class="price-unit">百万输出 ≈ ¥60</div>
|
||||||
|
</div>
|
||||||
|
<div class="price-card">
|
||||||
|
<div class="model-name">Claude Opus 4.6</div>
|
||||||
|
<div class="company">Anthropic</div>
|
||||||
|
<div class="price-label">百万输入 Token</div>
|
||||||
|
<div class="price-value">¥15</div>
|
||||||
|
<div class="price-unit">百万输出 ≈ ¥75</div>
|
||||||
|
</div>
|
||||||
|
<div class="price-card">
|
||||||
|
<div class="model-name">Gemini 3 Pro</div>
|
||||||
|
<div class="company">Google</div>
|
||||||
|
<div class="price-label">百万输入 Token</div>
|
||||||
|
<div class="price-value">¥14.4</div>
|
||||||
|
<div class="price-unit">百万输出 ≈ ¥57.6</div>
|
||||||
|
</div>
|
||||||
|
<div class="price-card highlight">
|
||||||
|
<div class="model-name">Qwen3.5-Plus</div>
|
||||||
|
<div class="company">阿里</div>
|
||||||
|
<div class="price-label">百万输入 Token</div>
|
||||||
|
<div class="price-value">¥0.8</div>
|
||||||
|
<div class="price-unit">价格仅为 1/18</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
<div class="bottom">同等质量下,Qwen3.5-Plus 成本不到国际顶尖模型的 1/18</div>
|
||||||
|
</div>
|
||||||
|
</body>
|
||||||
|
</html>
|
||||||
61
articles/007/diagrams/use-cases.html
Normal file
@ -0,0 +1,61 @@
|
|||||||
|
<!DOCTYPE html>
|
||||||
|
<html lang="zh-CN">
|
||||||
|
<head>
|
||||||
|
<meta charset="UTF-8">
|
||||||
|
<style>
|
||||||
|
* { margin: 0; padding: 0; box-sizing: border-box; }
|
||||||
|
body { width: 1200px; height: 680px; background: linear-gradient(135deg, #0f0c29, #302b63, #24243e); font-family: -apple-system, "PingFang SC", "Microsoft YaHei", sans-serif; display: flex; align-items: center; justify-content: center; color: #fff; }
|
||||||
|
.container { width: 1080px; }
|
||||||
|
h2 { text-align: center; font-size: 32px; margin-bottom: 44px; background: linear-gradient(90deg, #a78bfa, #60a5fa); -webkit-background-clip: text; -webkit-text-fill-color: transparent; }
|
||||||
|
.grid { display: grid; grid-template-columns: repeat(3, 1fr); gap: 20px; }
|
||||||
|
.card { background: rgba(255,255,255,0.06); border: 1px solid rgba(255,255,255,0.1); border-radius: 16px; padding: 28px 24px; text-align: center; }
|
||||||
|
.card .icon { font-size: 40px; margin-bottom: 12px; }
|
||||||
|
.card .title { font-size: 17px; font-weight: 700; color: #c4b5fd; margin-bottom: 8px; }
|
||||||
|
.card .model { font-size: 13px; color: #60a5fa; background: rgba(96,165,250,0.1); padding: 3px 12px; border-radius: 20px; display: inline-block; margin-bottom: 10px; }
|
||||||
|
.card .desc { font-size: 13px; color: rgba(255,255,255,0.6); line-height: 1.6; }
|
||||||
|
</style>
|
||||||
|
</head>
|
||||||
|
<body>
|
||||||
|
<div class="container">
|
||||||
|
<h2>Qwen3.5 使用场景推荐</h2>
|
||||||
|
<div class="grid">
|
||||||
|
<div class="card">
|
||||||
|
<div class="icon">💬</div>
|
||||||
|
<div class="title">日常对话 & 写作</div>
|
||||||
|
<div class="model">Qwen3.5-Plus</div>
|
||||||
|
<div class="desc">性价比最高,适合日常聊天、文案创作、知识问答</div>
|
||||||
|
</div>
|
||||||
|
<div class="card">
|
||||||
|
<div class="icon">💻</div>
|
||||||
|
<div class="title">编程辅助</div>
|
||||||
|
<div class="model">Qwen3.5-Max</div>
|
||||||
|
<div class="desc">代码能力强,支持多语言编程、代码审查、调试</div>
|
||||||
|
</div>
|
||||||
|
<div class="card">
|
||||||
|
<div class="icon">🖼️</div>
|
||||||
|
<div class="title">图片 / 文档分析</div>
|
||||||
|
<div class="model">Qwen3.5 多模态</div>
|
||||||
|
<div class="desc">原生视觉理解,读文档、看图表、识别 UI</div>
|
||||||
|
</div>
|
||||||
|
<div class="card">
|
||||||
|
<div class="icon">🎬</div>
|
||||||
|
<div class="title">视频内容理解</div>
|
||||||
|
<div class="model">Qwen3.5 多模态</div>
|
||||||
|
<div class="desc">支持最长 2 小时视频直接输入分析</div>
|
||||||
|
</div>
|
||||||
|
<div class="card">
|
||||||
|
<div class="icon">🏢</div>
|
||||||
|
<div class="title">企业私有化部署</div>
|
||||||
|
<div class="model">Qwen3.5-27B 开源版</div>
|
||||||
|
<div class="desc">数据不出境,可微调定制,无 API 依赖</div>
|
||||||
|
</div>
|
||||||
|
<div class="card">
|
||||||
|
<div class="icon">📱</div>
|
||||||
|
<div class="title">手机端 AI 助手</div>
|
||||||
|
<div class="model">Qwen3.5-2B</div>
|
||||||
|
<div class="desc">轻量本地运行,离线可用,隐私安全</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
</div>
|
||||||
|
</body>
|
||||||
|
</html>
|
||||||
BIN
articles/007/lmarena-labs-top.png
Normal file
|
After Width: | Height: | Size: 194 KiB |
BIN
articles/007/lmarena-labs.png
Normal file
|
After Width: | Height: | Size: 210 KiB |
BIN
articles/007/lmarena-mechanism.png
Normal file
|
After Width: | Height: | Size: 216 KiB |
BIN
articles/007/lmarena-ranking-table.png
Normal file
|
After Width: | Height: | Size: 127 KiB |
BIN
articles/007/lmarena-ranking.png
Normal file
|
After Width: | Height: | Size: 232 KiB |
BIN
articles/007/lmarena-top.png
Normal file
|
After Width: | Height: | Size: 222 KiB |
BIN
articles/007/moe-architecture.png
Normal file
|
After Width: | Height: | Size: 333 KiB |
BIN
articles/007/pricing-comparison.png
Normal file
|
After Width: | Height: | Size: 273 KiB |
BIN
articles/007/qwen-research.png
Normal file
|
After Width: | Height: | Size: 226 KiB |
BIN
articles/007/qwen35-benchmark.png
Normal file
|
After Width: | Height: | Size: 573 KiB |
BIN
articles/007/qwen35-blog.png
Normal file
|
After Width: | Height: | Size: 573 KiB |
BIN
articles/007/use-cases.png
Normal file
|
After Width: | Height: | Size: 320 KiB |