10xAI Wiki

12分钟阅读

China AI Model Landscape — 2025-2026

Overview

This page focuses on the model landscape inside the broader China AI ecosystem: model families, benchmarks, licensing and cost. From 2025 to 2026, Chinese open-source models moved from follower status to genuine global competition, often approaching top Western proprietary models on coding benchmarks while costing 10-100x less.

Major Players

Alibaba — Qwen

Parent: Alibaba Cloud (Alibaba Group)

Qwen is China’s most comprehensive open-source AI model family, with releases spanning general models, coding-specific models, and multimodal variants.

ModelParametersArchitectureSWE-benchContextLicense
Qwen3405B total / 35B activeMoE~60%128KApache 2.0
Qwen3-Coder-480B480B total / 35B activeMoE67%256K–1MApache 2.0
Qwen2.5-CoderVarious sizesDense~58%128KApache 2.0
Qwen-VLMultimodalDense + Vision—128KApache 2.0

Key facts:

  • Open-source strategy: Apache 2.0 license for all Qwen models
  • Best coding performance among open-source models globally (67% SWE-bench Verified)
  • 358 programming languages, 119 human languages
  • Released Qwen Code CLI (forked from Gemini Code)
  • Compatible with Claude Code, Cline, and other agents
  • Cost: ~0.10-0.80/M input tokens (vs. 15 for Claude Opus)

Team: Lin Junyang (P10) leads research. Lin Junyang announced departure March 2026, raising questions about future direction.

Moonshot AI — Kimi

Parent: Moonshot AI (founded 2023, backed by Alibaba, Tencent, Meituan)

Kimi K2 was released July 2025 and immediately disrupted the coding agent market:

ModelParametersSWE-benchPrice/M tokens
Kimi K21T total / 32B active65.8%$0.15 input

Key facts:

  • Native MCP support built-in
  • Kimi K2 × Claude Code integration guide published
  • Modified MIT license
  • Fastest-growing Chinese AI model by usage
  • Kimi K2 × OpenClaw integration available
  • Architecture uses DeepSeek-like innovations (mixture-of-experts, grouped-query attention)

Notable quote: “80% cost advantage over Claude 4 while matching performance on key benchmarks”

Zhipu AI — GLM

Parent: Zhipu AI (智谱AI), Tsinghua University spinoff

GLM has emerged as the most “agent-native” Chinese model:

ModelParametersSWE-benchTool-CallingLicense
GLM-4.5355B / 32B active64.2%90.6% successMIT
GLM-5744B / 40B active~69%~93%MIT
GLM-5-TurboQuantized variant~65%~90%MIT

Key facts:

  • Highest published tool-calling success rate of any open model (90.6%)
  • MIT license enables full commercial use
  • Runs on 8 H20 chips (Chinese data center GPU)
  • GLM-5 achieved 50 score on Artificial Analysis Intelligence Index v4.0 — first open-weight model to do so
  • “Agentic, 龙虾增强” — specifically optimized for agentic workflows

Key people: Luofuli (罗福莉) — leading RL researcher, joined Xiaomi AI Lab, driving the MiMo series

DeepSeek — V3 / V4 / R1

Parent: DeepSeek (Hangzhou, quantitative trading fund High-Flyer backed)

DeepSeek is the most disruptive force in the global AI landscape:

ModelContextKey AchievementPrice
DeepSeek V31M1M token context standard; Codeforces > GPT-5.4~1/7 of Claude Opus
DeepSeek V41MNear Claude Opus 4.6 non-thinking performanceVery low
DeepSeek R1128KFirst major Chinese reasoning modelLow

Defining moment: V4 Preview (April 2024) made 1M token context the standard across all tiers, previously exclusive to Google Gemini 1M.

Strategic significance: DeepSeek migrated from NVIDIA CUDA to Huawei Ascend chips, achieving 35x inference speed improvement. This breaks the assumption that frontier AI requires NVIDIA hardware — a major strategic development for China’s AI independence.

R2 delay: R2 delayed due to chip availability constraints (H100 export restrictions). Shows the real-world impact of hardware constraints.

ByteDance — Doubao / Seed

Parent: ByteDance (字节跳动)

ModelFocusNotable
Doubao-Seed-CodeCodingClaude Code integration available
Doubao Seed 2.0MultimodalVideo generation
Kimi K2Actually Kimi K2 = MoonshotNot ByteDance

ByteDance has the most aggressive AI integration into consumer products (Douyin/TikTok, Toutiao, Feishu).

Competitive Dynamics

Cost Competition

ModelInput Price/M tokensClaude Opus Ratio
Claude Opus 4.6$15.001x
GPT-5.4$2.500.17x
Kimi K2$0.150.01x
GLM-4.5$0.110.007x
Qwen3-Coder$0.100.007x
DeepSeek V4~$0.050.003x

Chinese models are 15-300x cheaper than Claude Opus.

Open-Source Strategy

All major Chinese labs have adopted aggressive open-source strategies:

  • Apache 2.0 (Qwen) or MIT (GLM) licenses
  • Weights publicly downloadable
  • Fine-tuning permitted
  • Commercial use allowed

This contrasts with Western proprietary models and creates ecosystem competition: Chinese models become the “Linux” of AI, Western models become the “Windows.”

Talent Wars

Key transfers in 2025-2026:

  • Luofuli left Zhipu → Xiaomi AI Lab → driving MiMo-V2 series
  • Lin Junyang (Qwen P10) announced departure March 2026
  • Multiple researchers moving between ByteDance, Alibaba, DeepSeek, Zhipu

[2026-07-20] Kimi K3：从”高性价比”到”接近前沿 + 资本市场冲击”

来源：Fable 5 × CursorBench / Kimi K3（2026-07-17）。

2026-07-17 Moonshot AI 发布 Kimi K3，标志中国模型从”高性价比编码模型”升级为”接近前沿的通用智能体模型”：

维度Kimi K2（2025-07）Kimi K3（2026-07）
总参数1T2.8T
上下文128K1M
智能指数（Artificial Analysis）—57（接近 GPT-5.5 / Claude Opus 4.8）
单任务成本—$0.94（约 Opus 4.8 一半）
前端代码竞技场—1679 分（第一）
API 定价$0.15 输入3 输入 / 15 输出（每百万 token）
权重Modified MIT完整权重最晚 2026-07-27

资本市场冲击：发布当晚催化全球 AI/半导体股抛售（摩根大通称”DeepSeek 2.0”，高盛称”去杠杆事件”）；港股智谱 −28.5%、MiniMax −15.6%，费城半导体指数较 6 月高点回落超 20% 入熊市。核心意义：打破了”美国前沿闭源模型享有稳定技术溢价”的预期。但注意 K3 定价较 K2 大幅上升（输入 0.15→3），且 2.8T 参数本地部署门槛极高——详见 Kimi K3 与 Scaling Law 之争。

[2026-08-06] 字节的 5 万亿豪赌与 Coding 竞赛（晚点独家）

来源：晚点 LatePost 独家（2026-08-06；内部信息均为转述，未经官方证实）。

ByteDance — Seed 最新动态：语言模型追赶、多模态领先、押注规模。

  • 讨论训练超 5 万亿参数模型（超 Qwen 3.8-Max 2.4T、Kimi K3 2.8T，国内已知最大），早期讨论阶段；由 Seed Foundation 负责人项亮主导、LLM 预训练数据负责人沈科配合。
  • Seed 2.0（2026-02，吴永辉执掌 Seed 后首个关键模型）市场反响有限——对照 GLM-5（被视为国内首个比肩 Opus 系列的模型）与 Kimi K3（多项第三方评测认为已接近海外闭源旗舰）。
  • Seedance 2.0（首个完整 MoE 视频模型、2000 亿参数）为全球最强视频模型，是火山引擎 MaaS 营收基本盘；但二季度以来 token 消耗与收入增速放缓。
  • 张一鸣定调（据晚点）：可接受短期落后、追求智能上限；反对蒸馏（本质是复制 Claude 能力）；别被 Coding 短期热点牵着走。

Coding 竞赛格局（2026 上半年）：

厂商进展（据报道）
AnthropicOpus 系列 Coding 打开程序员与 B 端市场，ARR 逼近并反超 OpenAI
智谱ARR 突破 10 亿美元
月之暗面ARR 突破 3 亿美元
字节跳动高薪引入 DeepSeek 核心研究员郭达雅负责 Coding 专项训练
阿里千问基模团队同期投入更多资源专攻 Coding 训练

商业化数据（据报道）：火山引擎为国内最大模型 API 卖家（份额约一半），2025 年收入约 150 亿元、2026 年内部目标超 400 亿元；豆包超 2 亿日活；豆包大模型 token 消耗 3 月 120 万亿 / 6 月 180 万亿，低于 250–300 万亿计划目标。

⚠️ 矛盾 [2026-08-06]：Seed 2.0 反响——wiki 源页（2026-03-03，“跨越鸿沟”）与本文（“反响有限”）评价相反；豆包 DAU——wiki 旧记录为峰值 1.45 亿，本文称超 2 亿。

？ OpenAI / Anthropic “传出”训练更大尺寸模型、xAI 下一代 6 万亿参数，均未获官方证实。

Sources: Claude Code + Doubao, Qwen3 Coder Debut, Chinese Models Comparison, DeepSeek V4 Six Things

关系图谱

反向链接