The HaxiTAG Forge team has surveyed all domestic tools, real prices, authoritative benchmarks, and the complete workflow from idea to launch. Targeting developers in mainland China, focusing on domestic tools and models.
Introduction: A Parallel Software Revolution
China's AI coding revolution is happening in sync with the global wave, but with an entirely different script.
While Cursor blazed past $2 billion ARR at the fastest pace in B2B SaaS history in Silicon Valley, two more structurally significant things happened in the Chinese market:
First, giants have thoroughly crushed the survival space for startups. ByteDance, Alibaba, Tencent, Baidu, and Huawei — five major players — simultaneously entered the market, offering high-quality AI coding tools for free or at extremely low prices. Globally, Cursor is a startup unicorn taking on GitHub Copilot; in the Chinese market, there are almost no startups of comparable scale — not because nobody wants to build them, but because the free strategy of the big players makes startup business models unviable from the very start.
Second, China invented the "Coding Plan" subscription model. In early 2026, as AI coding tools rapidly spread among domestic developers, the Coding Plan subscription model — replacing per-token billing with a fixed monthly fee — became a strategic high ground that domestic cloud vendors and LLM companies rushed to occupy. From Zhipu AI launching the GLM Coding Plan at the end of 2025, to Alibaba Cloud Bailian exploding the market with a rock-bottom price of ¥7.9/month in late February 2026, followed by Tencent Cloud in early March, a full-scale price war around AI coding subscriptions has broken out. This model targets the "bill shock" caused by per-token billing in agentic coding scenarios — a single complex task can trigger dozens of model calls and easily consume millions of tokens.
Here are the market numbers as of May 2026:
- According to EO Intelligence, China's AI code generation market reached RMB 6.5 billion in 2023 and is projected to grow to RMB 33 billion by 2028, a CAGR of 38.4%.
- According to IDC and QYResearch, China's AI coding tools market reached RMB 2.45 billion (2025), up 187.3% year-on-year, showing a pattern of "one superpower, two strong players, and multiple poles."
- To date, TRAE has over 6 million registered users, covering nearly 200 countries and regions; monthly active users exceed 1.6 million, spread across China, the US, Brazil, India, Japan, and others.
- In open-source self-hosted models, MiniMax M2.5 achieved an 80.2% score on SWE-bench, aligning with closed-source frontier models — Chinese models are closing in on the global top tier.
- The traditional per-token billing model for LLM APIs has revealed severe "bill shock" in AI coding agent scenarios, making the Coding Plan subscription model a strategic focus for domestic cloud vendors and LLM companies.
The tools are here. Prices are low enough. The network is no longer an issue.
The only question is: Which one is best for you?
Chapter 1: Complete Analysis of China's AI Coding Tools
Tools in the Chinese market are also divided into three categories, but the classification logic differs from the international market:
AI-Native IDE — Complete development environment with AI as first-class citizen. Examples: Trae, Qoder, CodeBuddy IDE, Tongyi Lingma IDE
AI Code Plugin — Enhancement plugins for VS Code / JetBrains. Examples: Tongyi Lingma plugin, Wenxin Kuaima, CodeGeeX
Open Source Terminal Agent — Bring your own API key, pay-as-you-go, fully controllable. Examples: OpenCode, iFlow CLI, Cline (using domestic models)
Category 1: AI-Native IDEs
TRAE CN (ByteDance)
Position: Absolute leader in China's AI-native IDE market with 41.2% market share, most comprehensive features, most aggressive updates.
Product Form:
TRAE China edition and International edition operate separately with significant differences:
| Dimension | TRAE CN (China edition) | TRAE (International edition) |
|---|---|---|
| Built-in models | Doubao 1.6 series (free), DeepSeek, Kimi | GPT-4o, Claude 3.7, etc. |
| Compliance | Domestic data localization, passed security review | No special compliance |
| Enterprise edition | Yes (includes private deployment) | Yes |
| Update frequency | Synchronized with intl edition, 100+ times/year | — |
Real Pricing:
| Plan | Monthly Price | Core Benefits | Notes |
|---|---|---|---|
| Individual Free | ¥0 | Free access to Doubao models, basic AI features | No usage limit (Doubao models) |
| Individual Pro | ~¥49/month | Advanced models (DeepSeek, Kimi, etc.), priority queue | Includes larger context |
| Enterprise | Contact sales | Private deployment, code security, productivity tracking | Minimum 100 users |
Deep-dive of Product Highlights:
TRAE continuously innovates in agent algorithms, achieving top scores on SWE-Bench Verified for both closed-source SOTA models and ByteDance's self-developed models.
TRAE's SOLO Mode is currently the closest domestic tool to an "agent engineering" form. From beta to general availability, the international edition's SOLO Mode reached a cumulative user penetration rate of 44% and a 1300% increase in queries; after the China edition's SOLO Mode launched, 3 out of every 10 TRAE developers adopted SOLO.
Cue Completion: TRAE's code completion feature Cue leads the industry in code continuation, predictive jumps, and cross-file code prediction. Over 50% of users actively use Cue daily for code editing. Since launch, Cue has recommended nearly 1 billion lines of code, with acceptance rates up over 80%.
Scale Proof: TRAE's average daily token consumption has increased nearly 700% in the last six months, generating a total of nearly 100 billion lines of code for global users in one year.
Practical Limitations:
- Doubao's self-developed models are stronger than open-source models in the same tier but weaker than Claude Sonnet series for extremely complex system architecture.
- SOLO Mode stability decreases under poor network conditions.
- Enterprise private deployment costs are still relatively high.
- China edition has fewer model switching options than the international edition (international edition can access GPT-5, etc.).
Best for: First choice for full-time domestic developers; rapid MVP for small to medium teams; WeChat Mini Program / Alipay Mini Program developers; developers with deep needs for domestic frameworks (Taro, Ant Design Pro).
Qoder (Alibaba, global-facing)
Position: Alibaba's agentic coding platform for the international market, benchmarked against Cursor, the most "global-ambitious" domestic tool.
Qoder was officially announced globally by Alibaba on August 22, 2025, as an Agentic intelligent coding platform. Its core concept evolves from a "code completion assistant" to an "AI engineer capable of autonomously completing complex tasks." It benchmarks Cursor, but Qoder's ambition is clearly larger.
Real Pricing:
Roughly half the price of Cursor. Business model: closed source, free + individual paid + enterprise paid.
| Plan | Monthly Price | Notes |
|---|---|---|
| Free | $0 | Basic agent capabilities, rate-limited |
| Individual | ~$10/month | About half the price of Cursor Pro |
| Enterprise | Contact sales | Qwen3-Max unlimited, private deployment |
Built-in Models: Qwen3-Max as main ( accessible to global users ), with Qwen3-Coder for coding tasks.
Relationship with Tongyi Lingma: Tongyi Lingma targets the domestic market, Qoder targets overseas — both are parallel products under Alibaba serving different audiences.
Practical Limitations:
- Domestic user experience is not as good as TRAE CN (optimized primarily for overseas).
- Agent mode understanding of Chinese projects slightly inferior to purely domestic-optimized versions.
- Ecosystem still rapidly catching up.
Tongyi Lingma IDE (Alibaba Cloud, domestic-focused)
Position: Mainstream product in the domestic enterprise market, compliance-first, deeply integrated with Alibaba Cloud ecosystem.
Tongyi Lingma is built by Alibaba Cloud based on the Qwen LLM series, with IDE plugin as its core form. It supports line-level completion (<300ms), MCP tool calls, natural language function generation, automatic unit test generation (target coverage 80%+), enhanced vulnerability detection, VPC private network calls, log auditing, and more.
Real Pricing:
| Plan | Monthly Price | Notes |
|---|---|---|
| Individual Free | ¥0 | 2,000 free requests per day |
| Individual Pro | ¥0 (time-limited free) | Advanced model features |
| Enterprise | from ¥79/month | MLPS Level 3 certification, private deployment |
Enterprise Competitiveness: Tongyi Lingma beat Baidu, Huawei, and Zhipu to win the Industrial and Commercial Bank of China's “Smart R&D Platform” project with a bid of RMB 2.11 million — a landmark event for AI coding entering financial-grade applications in China.
Suitable for: Developers in government/finance sectors with MLPS compliance needs; heavy Alibaba Cloud users; Java tech stack enterprise teams.
Tencent Cloud CodeBuddy
Position: Tencent's full-stack AI development platform, with Craft Agent as a key differentiating feature, deeply integrated with Tencent Cloud ecosystem.
Real Pricing:
| Plan | Monthly Price | Notes |
|---|---|---|
| Individual Free | ¥0 | Time-limited free for individual developers |
| Enterprise | ¥0 (time-limited free to activate) | ~¥158/month for min. 100 users after promo |
| Flagship (CouldPlay) | Contact sales | Private deployment, highest-level support |
Core Capabilities: Tencent Cloud CodeBuddy is the most feature-complete, supporting intelligent code completion (200+ languages), Craft Agent collaborative development, code review, test generation, and MCP protocol ecosystem integration.
Built-in Models: Hunyuan 2.0 series, with optional access to GLM-5, Kimi K2.5, MiniMax M2.5.
Integration with OpenClaw: Tencent Cloud Coding Plan provides both OpenAI-compatible and Anthropic-compatible access methods, supporting OpenClaw, CodeBuddy (Tencent self-developed), Claude Code, Cline, Cursor, OpenCode, CodeX, Kilo CLI, etc.
Practical Limitations:
- Hunyuan's coding capability ranks mid-tier among domestic models, weaker than DeepSeek V4 and Kimi K2.6.
- Enterprise actual pricing is opaque after the "time-limited free" period.
- IDE mode is less mature than TRAE.
Suitable for: WeChat/Mini Program ecosystem developers; heavy Tencent Cloud users; teams needing cloud-native deployment capabilities.
Baidu Wenxin Kuaima (Comate) IDE
Position: China's first multimodal, multi-agent collaborative AI IDE, Baidu's flagship product in the AI coding赛道.
Real Pricing:
| Plan | Monthly Price | Notes |
|---|---|---|
| Individual Standard | ¥0 | Basic features free |
| Individual Pro | ¥59/month | Advanced code intelligence, multi-agent collaboration |
| Enterprise | Contact sales | Private deployment, MLPS compliance |
Built-in Models: Wenxin 4.5 series (Baidu self-developed), optional access to DeepSeek V3/V4, Qwen series.
Differentiated Features: Multi-agent matrix collaboration (multiple AI agents handling frontend, backend, testing tasks separately); deep understanding of domestic documentation and SDKs (e.g., Baidu Maps SDK, PaddlePaddle).
Practical Limitations:
- Wenxin models generally weaker than DeepSeek V4 for code generation, stronger than same-tier open-source models.
- Multi-agent collaboration stability needs improvement for complex tasks.
- Limited optimization for non-Baidu tech stacks.
Huawei CodeArts (Pangu Code)
Position: Enterprise-grade AI coding platform for large state-owned enterprises, not for individual developers.
Key Differentiation: Huawei CodeArts supports project understanding at the scale of tens of millions of lines of code — the strongest capability among domestic tools for handling ultra-large-scale codebases, specifically serving large SOEs in power, aviation, banking with millions of lines of legacy code.
Price: Enterprise procurement only, no public pricing.
Suitable for: Teams developing with HarmonyOS; defense/government scenarios with high supply chain security requirements; financial institutions needing to work with ultra-large Java/C++ legacy codebases.
Category 2: AI Code Plugins
Tongyi Lingma (plugin form)
Available for VS Code and JetBrains suite, currently the most downloaded AI coding plugin in China. Free, supports VPC private network calls, suitable for enterprise intranet scenarios.
CodeGeeX (Tsinghua Zhipu AI)
Open source + free: Continuing the open style of Tongyi Lab, under Apache 2.0 license, fully open source; login via Qwen OAuth gives 2,000 free requests per day. It not only breaks down the ecosystem barriers of closed-source solutions but also is compatible with mainstream protocols like Anthropic, Google GenAI, OpenAI, not tied to a single model vendor. Technically backed by GLM series models, it is the plugin of choice for open-source contributors and academic researchers.
Category 3: Open Source Terminal Agents
OpenCode
Position: An open-source AI command-line tool based on Go that emerged in late 2025, quickly gaining recognition in the Chinese developer community with over 90k GitHub stars.
OpenCode was released in Q4 2025 by the SST/AnomalyCo team, built on the Bubble Tea TUI framework, supporting three forms: CLI, TUI, and IDE plugin. It is especially suitable for terminal power users (e.g., Neovim users).
Core Value for Chinese Developers: OpenCode is one of the most flexible CLI tools in terms of model support, supporting 75+ models, including almost all domestic APIs (DeepSeek, Qwen, Kimi, GLM, MiniMax, etc.), and is fully compatible with the OpenAI API protocol — meaning any domestic cloud vendor providing OpenAI-compatible interfaces can be connected.
Real Pricing: The tool itself is completely free and open source; costs come from your chosen API:
| API Combination | Average Monthly Cost (moderate use) |
|---|---|
| DeepSeek V4 Flash (official API) | ¥5-15 |
| Alibaba Cloud Bailian Coding Plan | ¥7.9 first month, then ¥39.9 |
| Qwen3-Max via Alibaba API | ¥10-30 |
| Local Ollama + Qwen3-32B | ¥0 (electricity only) |
Suitable for: Heavy Neovim/Vim users; independent developers with extreme cost requirements; senior engineers needing to customize agent workflows.
iFlow CLI
Position: Free open-source terminal AI coding tool based on Qwen3-Coder, focused on minimalism and zero cost.
Biggest highlight: completely free, with built-in domestic open-source models like GLM-4.6, MiniMax M2, capable of handling simple tasks, but code quality for complex projects needs improvement.
Honest assessment: iFlow positions itself as a "completely free entry-level tool" suitable for beginners just getting started with AI coding, not as a primary production tool.
OpenClaw — Special Mention
OpenClaw exploded in popularity in China's tech circles in early 2026, but its positioning is fundamentally different from the above tools:
OpenClaw was born in November 2025, initially just a "weekend project" by Peter Steinberger, a retired Austrian programmer. This open-source agent framework's core capability is simple: give LLMs local OS permissions, autonomously execute shell commands, directly take over your personal computer, schedule tasks across software, write code, and organize files.
OpenClaw is not a dedicated AI coding tool, but a general-purpose desktop AI agent that includes code generation capability (code_writer Skill). In China, Tencent, ByteDance, and Ant Group's Alipay have officially announced OpenClaw-related services, and the National Supercomputing Internet announced that its OpenClaw service has connected Feishu and WeChat Work.
How to use OpenClaw for coding tasks: Connect to domestic models (DeepSeek V4 or Kimi K2.5 recommended) via domestic Coding Plan, perform code generation and execution locally.
Important Security Warning: In February 2026, just three days after OpenClaw surpassed 100k stars, security researchers discovered 341 malicious "skills" in the ClawHub skill marketplace, accounting for 11.3% of the entire market, specifically stealing cryptocurrency wallets, account credentials, and system access. By early March, malicious skills had grown to 800+, about 20% of the market, and the severe remote code execution vulnerability CVE-2026-25253 was also officially disclosed.
Conclusion: OpenClaw is an interesting agent framework, but until security mechanisms are improved, it is not recommended for production-level coding on development machines containing codebases, API keys, or sensitive data. If you use it, run it in an isolated virtual machine.
Chapter 2: Complete Analysis of Domestic Coding Model Benchmarks
Domestic models officially entered the top tier in 2026 — but you need to read the details behind the numbers.
Domestic Coding Model Landscape (May 2026)
DeepSeek V4 Pro (Max) currently leads the BenchLM Chinese rankings with a score of 87; Kimi K2.6 follows closely at 84; GLM-5 (reasoning mode) and GLM-5.1 are both at 83; Qwen3.5 397B (reasoning mode) is at 79. The competitive landscape of China's frontier models is much more complex than the old GLM vs. Qwen vs. DeepSeek framework.
Detailed Benchmark Explanation
SWE-bench Verified (the most important coding benchmark)
What it measures: Ability to solve real GitHub Issues (500 validated real bug fixes).
2026 Domestic Model Scores:
| Model | Organization | SWE-bench Score | Open Source | Directly accessible in mainland |
|---|---|---|---|---|
| DeepSeek V4 Pro | DeepSeek | 81% | Open weights | ✅ Direct API |
| MiniMax M2.5 | MiniMax | 80.2% | Open | ✅ Direct API |
| MiniMax M2.7 | MiniMax | ~78% | Open | ✅ Direct API |
| GLM-5 (reasoning) | Zhipu AI | 77.8% | Partially open | ✅ Domestic API |
| GLM-5.1 | Zhipu AI | ~77% | Partially open | ✅ Domestic API |
| Kimi K2.5 (reasoning) | Moonshot | 76.8% | Open weights | ✅ Direct API |
| Kimi K2.6 | Moonshot | 65%+ (88.7 coding sub-score) | Open weights | ✅ Direct API |
| Qwen3-Coder | Alibaba | 70.6% | Open | ✅ Alibaba Cloud API |
| DeepSeek V4 Flash | DeepSeek | Close to V4 Pro | Open weights | ✅ Direct API |
| Doubao Doubao-Seed | ByteDance | Built into TRAE, not independently released | Closed | ✅ Volcano Ark API |
| Hunyuan HY 2.0 | Tencent | Not released | Closed | ✅ Tencent Cloud API |
Important Note: SWE-bench scores are heavily dependent on the test scaffold. Claude Opus 4.6 + Claude Code scores differ significantly from Claude Opus 4.6 + custom scaffold. OpenAI has also raised concerns about test data contamination. SWE-bench Pro (multilingual, standardized scaffold) is becoming a more reliable successor benchmark.
BenchLM Integrated Coding Sub-score
Moonshot's Kimi K2.6 scored 88.7 on the coding sub-score, the highest coding-specific score among Chinese models today, and the core driver of its 84 overall score.
Practical Benchmark: Complete Rails App Building Test
This is a real-world evaluation conducted by an independent researcher using a uniform prompt (build a complete Rails + RubyLLM + Hotwire + Docker application) on domestic models. The results are closer to real development experience:
Five Chinese models reached Tier B (usable, needs 1-2 hours of patching): K2.5, V4 Flash, MiMo, Qwen 3.6 Plus, GLM 5. The rest are not yet usable in production environments. The statement "China has caught up to the West in coding LLMs" needs a footnote: perhaps in synthetic reasoning benchmarks; but on the benchmark of delivering a complete application where every part runs, one model (Kimi K2.6) has caught up, and the rest still lag a generation behind.
This conclusion is very important: Benchmark scores and "ability to build a complete production-grade application" are two different things.
Domestic Model Selection Guide
Recommended by use case:
| Scenario | First Choice | Second Choice | Rationale |
|---|---|---|---|
| Bug fixing & code optimization | DeepSeek V4 Pro | Kimi K2.6 | Strongest SWE-bench performance |
| New complex project development | Kimi K2.6 | DeepSeek V4 Pro | Highest coding sub-score, strong complete app building |
| Code completion (high frequency, low latency) | DeepSeek V4 Flash | Qwen3-Coder | Low latency, extremely low cost |
| Chinese doc understanding + coding | Qwen3-Max / Qwen3-Coder | GLM-5.1 | Best understanding of Chinese technical docs |
| Open source self-hosted (local) | Qwen2.5-Coder 32B | DeepSeek-Coder-V2 | Can run locally, sufficient performance |
| Math/algorithm tasks | DeepSeek R1 | Kimi K2.5 (reasoning mode) | Strongest reasoning chain |
| Cost priority | DeepSeek V4 Flash | Qwen3-Coder (free tier) | Lowest cost per million tokens |
What a 5-point gap means in practice
Taking SWE-bench as an example, a 5-point gap = solving 1 more task out of every 20 real engineering tasks. However, the gap is amplified in the following scenarios:
- Legacy codebases (>100k lines): Differences in context understanding are sharply amplified here. DeepSeek V4 Pro's 128K context (extremely low cost after caching) gives it a clear advantage over Qwen3-Coder's 256K context (though Qwen's long context has better cost-effectiveness).
- Codebases with Chinese comments and documentation: Here, Qwen's optimization for Chinese understanding can compensate for a 1-2 point SWE-bench gap.
- Simple single-file tasks: The gap between almost all mainstream models is negligible; cost becomes the deciding factor.
Chapter 3: Three-Question Decision Framework for Chinese Developers
Replace endless parameter comparisons with three questions.
Question 1: What are your network environment and data requirements?
A. Can access overseas APIs, no special data compliance requirements → You can use the full range of tools, refer to the international version guide combined with the Coding Plan section in this guide. → Recommendation: Cursor Pro + DeepSeek V4 API, or TRAE international edition.
B. Can only use domestically accessible tools and models (mainland development environment) → Proceed to Question 2.
C. Has MLPS / industry compliance requirements (finance/government/healthcare) → Choose directly: Tongyi Lingma Enterprise Edition (MLPS Level 3 certified, most widely deployed in SOEs) or Huawei CodeArts (highest compliance level).
Question 2 (only for Category B users): Do you plan to write code?
A. No, or have no programming foundation at all
| Project Type | Recommended Tool | Rationale |
|---|---|---|
| Complete web app (backend + frontend) | TRAE CN free + SOLO Mode | Closest experience to Lovable among domestic tools |
| WeChat Mini Program / internal corporate tools | TRAE CN or CodeBuddy | Built-in support for WeChat/Tencent ecosystem |
| Data dashboard | Tongyi Lingma IDE | Deep integration with Alibaba data products (MaxCompute, DataWorks) |
| Validate product idea (fastest) | TRAE CN SOLO Mode | Currently the closest to fully automated build experience |
B. Can write code, or can read code → Proceed to Question 3.
Question 3 (only for Category B users who can write code): What is your core need?
| Need | Recommended Solution | Average Monthly Cost |
|---|---|---|
| Best overall experience, primary IDE | TRAE CN Individual | ¥0-49 |
| Code quality first, accept command line | OpenCode + DeepSeek V4 Pro API | ¥30-80 (API fees) |
| Extreme cost optimization, strong technical skills | OpenCode + Alibaba Cloud Bailian Coding Plan | ¥7.9 first month, then ¥39.9 |
| Enterprise team with compliance requirements | Tongyi Lingma Enterprise | from ¥79/person/month |
| Tencent Cloud ecosystem lock-in | CodeBuddy Enterprise | Time-limited free, then reference ¥158/month |
| Open source enthusiast, full control | OpenCode + local Ollama | ¥0 (local compute) |
Quick Decision Tree for Chinese Developers
Network/compliance constraints?
├── Has MLPS compliance → Tongyi Lingma Enterprise / Huawei CodeArts
│
├── Domestic tools & models only
│ ├── No coding → TRAE CN (SOLO Mode)
│ └── Can code
│ ├── Best IDE experience → TRAE CN Individual
│ ├── Highest code quality → OpenCode + DeepSeek V4 Pro
│ └── Lowest cost → OpenCode + Alibaba Cloud Bailian Coding Plan
│
└── Can use overseas tools
├── Full-time developer → Cursor Pro + domestic Coding Plan (hybrid)
└── Pursuing ultimate quality → Claude Code + Max subscription (requires overseas payment)
Chapter 4: Complete Domestic Pricing Breakdown
A major innovation in the 2026 Chinese market: the Coding Plan subscription model.
What is a Coding Plan?
Because AI developers consume large amounts of tokens, major domestic LLM providers have launched Coding Plans (coding-dedicated subscription packages) that provide high token allowances, dedicated code models, and high-speed invocation rights, specifically to support the stable operation of AI coding agents. By trading a fixed monthly fee for ample model invocation quotas, AI developers can completely escape token-based billing anxiety.
Why is the Coding Plan important? A single complex code refactoring using agent mode may trigger 30-50 API calls, consuming 2-5 million tokens. Under pay-as-you-go billing, the cost of a single task could be as high as ¥30-80; the Coding Plan turns that cost into ¥7.9 for an entire month.
Horizontal Comparison of Domestic Mainstream Coding Plans (May 2026)
From Zhipu AI first launching the GLM Coding Plan at the end of 2025, to Alibaba Cloud Bailian exploding the market with a rock-bottom price of ¥7.9/month in late February 2026, followed by Tencent Cloud in early March, a full-scale price war around AI coding subscriptions is on.
| Platform | First Month Price | Standard Monthly Price | Supported Main Models | Tool Compatibility |
|---|---|---|---|---|
| Alibaba Cloud Bailian | ¥7.9 | ¥39.9 (Lite) / ¥199 (Pro) | Qwen3.5+, Qwen3-Coder, MiniMax M2.5, GLM-5, Kimi K2.5 | OpenCode, Cline, Cursor, Claude Code, Qwen Code |
| Tencent Cloud | ¥7.9 | ¥39.9 (Lite) / ¥199 (Pro) | Hunyuan HY 2.0, GLM-5, Kimi K2.5, MiniMax M2.5 | OpenClaw, CodeBuddy, Cline, Cursor, OpenCode |
| Volcano Ark (ByteDance) | — | ¥49 (Lite) / ¥400 (Max) | Doubao Seed 2.0, MiniMax M2.5, Kimi K2.5, GLM-4.7, DeepSeek V3.2 | Claude Code, Cursor, OpenCode, Cline |
| Baidu Qianfan | ¥7.9 | ¥40 (Lite) / ¥200 (Pro) | GLM-5, Kimi K2.5, MiniMax M2.1, DeepSeek V3.2 | OpenClaw, Cline, Cursor |
| MiniMax Official | — | ¥39 (Basic) / ¥199 (Pro) | MiniMax M2.5-highspeed | Mainstream tools |
| Kimi Official | — | ¥39.9 (Lite) / ¥199 (Pro) | Kimi K2.5 | Mainstream tools |
| UCloud | ¥6.9 | ¥39 (Lite) / ¥199 (Pro) | DeepSeek, GLM, Kimi, MiniMax | Mainstream tools |
Alibaba Cloud Bailian's core advantage: Bailian supports the largest number of models, including Qwen3.5-Plus, Qwen3-Max, Qwen3-Coder-Next, Qwen3-Coder-Plus, MiniMax M2.5, GLM-5, Kimi K2.5, GLM-4.7 — 8 mainstream models, the highest model diversity among platforms. Supported tools cover Qwen Code, OpenClaw, OpenCode, Claude Code, Codex, Cline, Cursor, Kilo CLI, and over 10 other mainstream AI coding tools.
Volcano Ark's key difference: Volcano Ark measures by "requests" rather than "prompts"; one user question typically triggers multiple model calls, each consuming one request quota. It supports mainstream tools like Claude Code, Cursor, OpenCode, Cline, and additionally offers a Max Plan (¥400/month, quarterly payment ¥1,200).
Direct API Pricing (for pay-as-you-go scenarios)
DeepSeek API:
DeepSeek V4 is the newest flagship model, released in early March 2026, with input price of $0.30 per million tokens and output price of $0.50 per million tokens. After cache hit, the input price is only $0.03 per million tokens, a 90% reduction. This means that for production applications with a common prefix (system instructions, tool definitions, document templates), the effective input cost can be as low as $0.05/M tokens.
| Model | Input (cache miss) | Input (cache hit) | Output | RMB conversion (reference) |
|---|---|---|---|---|
| DeepSeek V4 Pro | $0.30/M | $0.03/M | $0.50/M | ~¥2/M output |
| DeepSeek V4 Flash | $0.14/M | $0.0028/M | $0.28/M | ~¥1/M output |
| DeepSeek R1 | $0.55/M | $0.14/M | $2.19/M | ~¥16/M output |
Note: DeepSeek V4 Pro currently receives a 75% discount, valid until May 31, 2026.
Domestic Main Model API Pricing (RMB):
| Model | Provider | Input Price | Output Price | Notes |
|---|---|---|---|---|
| Qwen3-Max | Alibaba | ¥0.8/M | ¥3.2/M | Supports 1M context |
| Qwen3-Coder | Alibaba | ¥0.4/M | ¥1.6/M | Optimized for code |
| GLM-5 | Zhipu | ¥1.2/M | ¥4.8/M | Reasoning-enhanced version more expensive |
| Kimi K2.5 | Moonshot | ¥0.6/M | ¥2.4/M | 256K context |
| MiniMax M2.5 | MiniMax | ¥0.5/M | ¥2.0/M | 1M context |
| Doubao Seed-2.0-pro | ByteDance | ¥0.7/M | ¥2.8/M | Volcano Ark only |
| Hunyuan HY 2.0 | Tencent | ¥0.8/M | ¥3.2/M | Tencent Cloud |
Pricing advantage: From DeepSeek the "price butcher" with ¥0.2/million tokens, to Volcano Ark's "milli-pricing" era, developers have many cost-effective choices.
Comprehensive Cost Scenario Analysis
Scenario A: Student / hobbyist developer, 20 hours/month
- Plan: TRAE CN free + Alibaba Cloud Bailian Coding Plan (¥7.9 first month)
- Average monthly cost: ¥7.9-39.9
Scenario B: Full-time independent developer, 8 hours/day
- Plan: TRAE CN Individual (¥49) + DeepSeek V4 API for complex tasks (¥30-80)
- Average monthly cost: ¥80-130
Scenario C: Strong technical skills, extreme cost control
- Plan: OpenCode + Alibaba Cloud Bailian Coding Plan Pro (¥7.9 first month, then ¥199)
- Average monthly cost: ¥7.9 first month, then ¥199 (includes unlimited calls to 8 major models)
Scenario D: 5-person tech team, general compliance requirements
- Plan: TRAE CN Enterprise or CodeBuddy Enterprise
- Average monthly cost: ¥500-1000 (depends on scale and feature needs)
Scenario E: SOE/financial institution, MLPS compliance
- Plan: Tongyi Lingma Enterprise (from ¥79/person/month)
- Average monthly cost: from ¥79+/person, one-time additional cost for private deployment
Chapter 5: Complete Workflow from Idea to MVP for Chinese Developers
Designed specifically for mainland China development environment, including common domestic tech stacks and deployment options.
Stage 0: Idea Clarification (mandatory, 2-3 hours)
Before opening any AI tool, run the following conversation in TRAE's Chat mode or on Kimi.ai (directly accessible in China):
I want to build [your idea].
Please help me:
1. Describe the core value proposition in one sentence.
2. List the 3 core features of the MVP (minimize, don't be greedy).
3. Identify the 3 biggest technical risks.
4. Recommend the most suitable tech stack for the mainland China development environment.
Constraints:
- Need to deploy quickly (no more than 2 weeks)
- Primary users are in mainland China
- Need to integrate WeChat Pay / Alipay [if applicable]
- Team size: [number] people
China-specific tech stack considerations: Domestically developed SaaS products need to consider the WeChat ecosystem (Mini Programs, Official Accounts, WeChat Work) and compliance requirements (ICP filing, data localization). AI tool support for these scenarios varies significantly; TRAE CN and Tongyi Lingma understand the domestic ecosystem much better than Cursor.
Stage 1: Tech Stack Decision (domestic adaptation)
Non-technical founders (using AI-native IDE): Directly use TRAE CN SOLO Mode, no tech stack decision needed.
Developers with technical background:
| Project Type | Recommended Tech Stack | Domestic Deployment | AI Tool |
|---|---|---|---|
| SaaS web app (no special requirements) | Next.js + Supabase (or Tencent Cloud CDB) + Stripe/WeChat Pay | Alibaba Cloud ECS / Tencent Cloud CVM | TRAE CN + OpenCode |
| WeChat Mini Program | Taro + NestJS + MySQL | Tencent Cloud CloudBase | TRAE CN (best understanding of Mini Program frameworks) |
| Data tool | Python + FastAPI + PostgreSQL | Alibaba Cloud RDS | OpenCode + DeepSeek API |
| Internal enterprise system | Vue3 + Spring Boot + MySQL | Private server or hybrid cloud | Tongyi Lingma Enterprise |
| Mobile App | React Native / Flutter + NestJS | Tencent Cloud / Huawei Cloud | TRAE CN |
Preferred domestic deployment combo (best value for money):
Frontend: Vercel (accessible) or Cloudflare Pages
Backend: Railway or Tencent Cloud CloudBase
Database: Supabase (has SG node, acceptable latency in mainland) or PlanetScale
Object storage: Tencent Cloud COS or Alibaba Cloud OSS
Stage 2: Project Scaffolding
Initial prompt for TRAE CN (adapted for domestic scenarios):
I want to build an [application name] for mainland China users.
Tech stack:
- Frontend: Next.js 14 (App Router) + TypeScript + Tailwind CSS
- Backend: Next.js API Routes + Prisma ORM
- Database: PostgreSQL (Tencent Cloud CDB or Supabase)
- Authentication: NextAuth.js + phone number verification (using Tencent Cloud SMS or Alibaba Cloud SMS)
- Payment: WeChat Pay + Alipay
Core features (MVP):
1. [Feature 1]
2. [Feature 2]
3. [Feature 3]
Please help me:
1. Design a database schema that meets domestic regulatory requirements (including user phone number, real-name information fields)
2. Build the project directory structure
3. Create the phone number verification code login flow (domestic users prefer phone number over email)
4. Set up the basic framework for WeChat Pay
Note: Write code comments in Chinese, use English variable names.
Stage 3: Core Feature Development
Special considerations for domestic development:
WeChat ecosystem integration: This is one of the most complex technical challenges in domestic SaaS products. AI tool capability differences are extremely significant here.
# Test AI tool understanding of WeChat ecosystem
"Please implement WeChat web authorization login (OAuth 2.0),
including getting access_token, getting user openid and user information,
handling various error codes returned by WeChat,
and silent refresh logic after token expiration"
Good tool (TRAE CN): generates basically correct implementation, understands the difference between openid vs. unionid
Poor tool (unfamiliar with domestic ecosystem): generates code that confuses standard OAuth flow with WeChat-specific flow
Alipay/WeChat Pay integration:
"Please implement WeChat Pay V3 JSAPI payment,
including:
1. Unified order interface call
2. Payment signature generation (SHA256withRSA)
3. Payment callback signature verification
4. Order status query
Use the official wechatpay-node-v3 SDK"
For such tasks, TRAE CN or Tongyi Lingma are recommended; their understanding of domestic payment SDK documentation is significantly better than Cursor.
ICP filing related code: Remember to have AI generate the ICP filing information display module at the bottom of the website, a legal requirement for domestic websites:
"Add an ICP filing number display component at the bottom of all pages,
filing number: [your filing number]
link to: https://beian.miit.gov.cn/"
Stage 4: UI Polishing
Domestic design guidelines reference: Domestic user UI preferences differ from international markets. Have AI refer to the following design language:
"Please adjust the UI following these design principles:
- Moderate information density (refer to WeChat design guidelines)
- Primary color: [your brand color]
- Use Ant Design specifications (most familiar to domestic users)
- Ensure normal display in the WeChat built-in browser on iOS/Android
- Adapt to 375px (iPhone SE) and 414px (iPhone Pro Max)
- Moderate border radius (8-12px), in line with domestic aesthetics"
Stage 5: Testing
China-specific testing scenarios (easily overlooked):
| Testing Scenario | Description | Testing Method |
|---|---|---|
| WeChat built-in browser compatibility | WeChat WebView CSS/JS support differs from standard browsers | Use WeChat Developer Tools + real device |
| Weak network testing | Network quality in some domestic areas is unstable | Chrome DevTools network throttling |
| Domestic phone adaptation | System browser compatibility on Huawei/Xiaomi/OPPO | Real device testing or BrowserStack |
| User phone number verification | Ensure correct handling of +86 prefix | — |
| Third-party login | Fallback solution when WeChat web authorization is opened outside H5 | — |
# Have AI generate test cases for domestic scenarios
"Please generate test cases for the phone number SMS verification code login feature,
including:
1. Normal flow (get verification code → verify → login success)
2. Verification code expiration (retry after 5 minutes)
3. Account lock logic after 5+ incorrect verification code entries
4. Rate limiting for frequent verification code requests (same phone number can only send once per 60 seconds)
5. Phone number format validation (11 digits, starting with 1)"
Stage 6: Domestic Deployment
Deployment options accessible in China:
| Scenario | Frontend | Backend | Database | Estimated Monthly Cost |
|---|---|---|---|---|
| MVP validation (no ICP) | Vercel free / Cloudflare Pages | Supabase (Singapore node) | Supabase free | ¥0 (slower access speed) |
| Early product (ICPs filed) | Tencent Cloud CloudBase / Alibaba Cloud OSS + CDN | Tencent Cloud CVM / Alibaba Cloud ECS | Tencent Cloud CDB / RDS | ¥100-300 |
| Growth stage | Alibaba Cloud CDN + OSS | Multi-region ECS + SLB | PolarDB | ¥500+ |
| Mini Program product | WeChat Cloud Development | WeChat Cloud Function | WeChat Cloud Database | ¥0-200 (pay-as-you-go) |
Pre-deployment China-specific checklist:
"Please help me check the following domestic compliance deployment items:
1. Is the ICP filing number displayed at the bottom of all pages?
2. Does user registration collect phone numbers? (domestic real-name system requirement)
3. Are the privacy policy and user agreement pages complete?
4. Is there a privacy statement compliant with PIPL?
5. Is payment functionality integrated with a legitimate payment aggregator SDK?
6. Is HTTPS configured? (requirement for domestic app store listing)
7. If the Mini Program involves medical/financial content, are the corresponding qualifications stated?"
Chapter 6: Prompt Template Library Exclusively for Chinese Developers
The following templates are optimized for domestic tech stacks and development habits, usable in TRAE CN, Tongyi Lingma, and OpenCode.
Template 1: Domestic Project Scaffolding (including WeChat ecosystem)
You are a full-stack engineer familiar with the domestic technology ecosystem. I am building [project name] for mainland China users.
Tech stack:
- Frontend: Next.js 14 (App Router) + TypeScript + Tailwind CSS + Ant Design
- Backend: Next.js API Routes + Prisma ORM
- Database: PostgreSQL (Tencent Cloud CDB or PlanetScale)
- Authentication: Phone number SMS verification code (primary) + WeChat web authorization (backup)
- SMS service: Tencent Cloud SMS (most stable domestically)
- Payment: WeChat Pay V3 + Alipay (choose one or both)
- File storage: Tencent Cloud COS or Alibaba Cloud OSS
Core entities (adjust according to business):
- User: phone number, openid, realName, idCard (if real-name required), createdAt
- [Entity 2]: [fields]
MVP features (by priority, only list 3):
1. [Feature 1]
2. [Feature 2]
3. [Feature 3]
Please help me:
1. Design the file directory structure (including middleware layer for authentication and logging)
2. Design a PIPL-compliant Prisma schema (including soft delete fields)
3. Implement the complete phone number SMS verification code login flow
4. Build the basic framework for WeChat Pay V3 (including callback signature verification)
Write code comments in Chinese, use English variable names, in line with domestic team collaboration habits.
Template 2: WeChat Pay / Alipay Integration
I need to integrate [WeChat Pay V3 / Alipay] JSAPI payment functionality for [project name] (applicable to H5 web pages or Mini Programs).
Current environment:
- Node.js + Next.js App Router
- Existing SDK: [wechatpay-node-v3 / alipay-sdk]
- Database: PostgreSQL + Prisma
Need to implement:
1. Create order endpoint (/api/payment/create)
- Input: product ID, quantity, user openid
- Validation: user logged in, sufficient stock, correct amount
- Output: prepay_id and signature parameters required for JSAPI payment
2. Payment callback endpoint (/api/payment/notify)
- Verify WeChat/Alipay signature (prevent fake callbacks)
- Idempotent handling (only process callback once for same order number)
- Update order status and inventory
- Trigger business logic (send purchase success notification)
3. Order status query (/api/payment/query)
4. Refund endpoint (/api/payment/refund)
Please:
1. First explain your implementation approach and security considerations
2. Implement complete code, including error handling and logging
3. List test steps (including sandbox environment configuration)
4. Point out common payment pitfalls (e.g., openid vs. unionid, callback idempotency, reconciliation logic)
Template 3: Chinese Codebase Debugging
My [project name] has a bug, need your diagnosis.
Error message (full copy):
[error stack]
Reproduction steps:
1. [Step 1]
2. [Step 2]
Expected: [description]
Actual: [description]
Relevant files:
[paste key code]
Business context (important):
- This is a [type] feature for [scenario]
- Users encounter this issue when [trigger condition]
- Most recent code change: [description]
Please:
1. Analyze the root cause (2-3 possibilities)
2. Provide a fix (prioritize best practices for common domestic tech stacks)
3. Tell me how to verify the fix is successful
4. If it's a middleware/framework issue, suggest whether there is a more mature domestic alternative
Template 4: Domestic Compliance Security Review
Please conduct a security review of the following code, focusing on:
General security items:
1. SQL injection risk (all database queries)
2. XSS vulnerabilities (handling and rendering of user input)
3. Authentication and authorization bypass (API route protection)
4. Exposure of sensitive data (logs, API responses, error messages)
5. Missing rate limiting (especially important for login endpoints, SMS sending endpoints)
China-specific compliance items:
6. PIPL compliance
- Is there clear purpose notice for data collection?
- Are phone numbers and real-name information encrypted in storage?
- Does it support user account deletion and data export/deletion?
7. Payment security
- Is the payment callback signature verified?
- Is the order amount recomputed on the server side (client-side amount input prohibited)?
8. SMS verification code security
- Is there anti-brute-force validation (frequency limits on same IP/phone number)?
- Is the verification code generated server-side (client-side code input prohibited)?
Code:
[paste code]
For each issue, please:
- Describe the vulnerability and attack scenario
- Provide fix code
- Label: High risk / Medium risk / Low risk
Template 5: Domestic Tech Stack Architecture Review
Please review my application architecture against the following dimensions:
Current architecture:
[describe tech stack, main modules]
Business stage: [MVP validation / growth / scale]
Current user count: [number] MAU
Target in 6 months: [number] MAU
Need ICP filing / MLPS certification: [Yes/No]
Please evaluate:
1. Scalability: When user count grows 10x, where will the bottlenecks be?
(Focus: domestic bandwidth costs, database connection pools, CDN configuration)
2. China-specific risks:
- Does it rely on overseas services with unstable access within China?
- Does data need localized storage (finance/healthcare industries)?
- Fallback plan for overseas CDNs in certain network environments?
3. Cost structure:
- Trade-off between domestic cloud (Tencent/Alibaba) vs. overseas cloud
- API costs: over-reliance on expensive models?
4. Technical debt:
- Which choices are convenient now but will cause trouble in the future?
- WeChat SDK version management and update strategy?
For each issue: analyze current situation → state when it becomes a real problem → provide step-by-step improvement path.
Template 6: TRAE Rule File (TRAE.md / .cursorrules for China)
# Project Specification File - Domestic Development Edition
## Project Overview
[Project Name] is a [type] product for mainland China users.
Core features: [2-3 sentence description]
## Tech Stack
- Frontend: Next.js 14 App Router + TypeScript + Tailwind CSS + Ant Design 5
- Backend: Next.js API Routes + Prisma ORM 5.x
- Database: PostgreSQL (Tencent Cloud CDB)
- Authentication: Phone number SMS verification code (Tencent Cloud SMS) + JWT
- Payment: WeChat Pay V3 (integrated)
- Object storage: Tencent Cloud COS
- Deployment: Tencent Cloud CloudBase (frontend) + CVM (backend)
## Coding Standards
### Naming Conventions
- React components: PascalCase (UserProfile.tsx)
- Utility functions: camelCase (formatPhoneNumber.ts)
- API routes: kebab-case directories (user-profile/route.ts)
- Database tables: snake_case (user_orders)
### Code Style
- Use TypeScript strict mode, no any
- Prefer async/await, avoid .then() chaining
- All async functions must have try/catch
- Prefer functional components with React Hooks
### China-Specific Standards
- Phone numbers: 11 digits, without country code (+86)
- Amount fields: unit is "fen" (integer), do not use floating point
- Logging: prohibited to log phone numbers, ID numbers, or other sensitive information
- Payment callbacks: must verify signature, do not trust amount from client
- All queries must verify current user's permissions (userId isolation)
### Database
- All database operations via Prisma, raw SQL prohibited
- All user tables must include createdAt, updatedAt, deletedAt fields (soft delete)
- Table operations involving financial data must use transactions
### API Security
- All API routes must verify authentication status
- Login endpoints (phone + verification code) must have rate limiting: same phone number once per 60 seconds
- SMS sending endpoints: add IP rate limiting: same IP max 5 times per 10 minutes
## Prohibited Items
- Do not hardcode AppID, AppSecret, AccessKey, or other secrets
- Do not log user phone numbers, passwords, or payment information
- Do not trust payment amounts from the client (recompute server-side)
- Do not skip WeChat Pay callback signature verification
- Do not store plaintext passwords or unencrypted sensitive information
## Domestic Ecosystem Hints
- User ID system: distinguish between openid (unique within Mini Program) and unionid (unique across products)
- WeChat Pay: unified order → get payment parameters → frontend invoke payment → backend receive callback → update order status
- SMS verification code: 6 digits, valid for 5 minutes, invalidated immediately after verification
## When Unsure
Ask the user instead of making assumptions. For decisions related to domestic compliance, state the uncertainty explicitly.
Chapter 7: 7 Unique Mistakes That Kill Chinese Developer Projects
These are the most common failure patterns in domestic AI coding projects, never mentioned in international guides.
Mistake 1: Generating WeChat Pay code with AI and going live without sandbox verification
Symptom: Payment works fine in testing (using mock data), but fails for real users after launch, or order status doesn't update after successful payment.
Root cause: The WeChat Pay sandbox environment differs subtly from production: signature algorithm details, callback IP whitelisting, certificate path configuration. AI tools tend to generate "looks correct" code that misses details.
Solution:
- Must apply for a sandbox account in the WeChat Pay merchant backend and test the complete payment flow in the real sandbox environment.
- Have AI specifically explain each step of the callback signature verification:
"Please explain this payment callback signature verification code line by line: what each step does, and why." - Use Tencent's official wechatpay-node-v3 SDK, not AI's own signature implementation.
Mistake 2: Ignoring ICP filing, leading to forced shutdown after launch
Symptom: Your product deploys and works normally on Tencent Cloud/Alibaba Cloud, then 2-4 weeks later suddenly becomes inaccessible; customer service says you need an ICP filing.
Root cause: Websites in mainland China must complete ICP filing with the Ministry of Industry and Information Technology. Cloud service providers are obligated to restrict access to unfiled websites. The process typically takes 7-20 business days, during which the website cannot operate normally.
Solution: Submit the ICP filing application in your cloud provider's backend at the same time you write the first line of code. Alternatively, use a non-mainland node (Hong Kong/Singapore) for temporary deployment during filing, then migrate to a mainland node after filing is complete.
Mistake 3: Data compliance issue: AI-generated registration form missing user agreement
Symptom: After launch, legal or regulatory authorities point out that the user registration flow lacks a clear privacy policy consent checkbox, violating the PIPL.
Solution: When asking AI to generate user registration functionality, must explicitly include:
"Add to the bottom of the registration form:
□ I have read and agree to the User Agreement and Privacy Policy (required)
And ensure:
1. The user must check the box to submit (server must also validate this field)
2. 'User Agreement' and 'Privacy Policy' have separate links to the full text
3. Record the timestamp and version number of user consent (stored in database)"
Mistake 4: Using overseas CDNs directly from AI-generated code in production
Symptom: Your product loads extremely fast on your development machine, but users report very slow page loads or failure to load certain resources.
Root cause: AI tools (especially international versions) often generate code that defaults to referencing overseas CDNs like unpkg.com, jsdelivr.net, cdnjs.cloudflare.com. These CDNs are extremely unstable to access from mainland China.
Solution: Have AI check and replace all external resource references:
"Please check all package.json dependencies and external resource URLs in HTML/JS,
replace all overseas CDNs (unpkg, jsdelivr, cdnjs.cloudflare.com) with:
- npm bundled locally (via webpack/Next.js bundling)
- or domestic CDNs (bootcdn.net, staticfile.org)
- Fonts: replace Google Fonts with local font files or domestic font services"
Mistake 5: AI-generated SMS verification code feature gets abused
Symptom: Your SMS bill suddenly spikes (thousands of RMB per day). Upon inspection, someone is using scripts to bulk request verification codes.
Root cause: AI-generated SMS verification code interfaces typically have only basic implementation, lacking anti-abuse mechanisms.
Solution:
"Add multi-layer anti-abuse protection to the SMS verification code sending endpoint:
1. Same phone number: can only send once per 60 seconds (Redis storage, key: sms:phone:{phone number})
2. Same IP: max 5 times per 10 minutes (Redis storage, key: sms:ip:{IP})
3. Same phone number: max 10 times per day (Redis, key: sms:daily:{phone number}:{date})
4. After 5 wrong verification code entries for the same phone number, lock that phone number for 30 minutes
All checks must be performed server-side, do not trust any parameters from the client."
Mistake 6: Using AI-generated .env variable names and pushing to a public Git repository
Symptom: Alibaba Cloud or Tencent Cloud security system sends a warning: AccessKey information detected in repository, account security at risk.
This mistake is mentioned in international guides, but has additional risks in the Chinese market: After a Tencent Cloud / Alibaba Cloud AccessKey is leaked, attackers can quickly launch large numbers of GPU instances for crypto mining, potentially causing tens of thousands of RMB in losses per day.
Solution: Use dedicated Secret management services (Tencent Cloud SSM or Alibaba Cloud KMS) instead of .env files for production secrets. Also enable "Secret Scanning" in your GitHub repository settings (free).
Mistake 7: Over-relying on AI-generated code and skipping real device testing
Symptom: Your product displays correctly in the mobile emulation of browser dev tools, but on real Huawei Mate/Xiaomi 14/OPPO phones, the layout is broken or some features don't work.
Root cause: Chinese Android phones' system browsers (highly customized Chromium variants) and the WeChat built-in browser (WebView) have subtle differences in CSS and JavaScript support compared to standard browsers. AI training data comes primarily from standard web environments and has limited understanding of these domestic differences.
Solution: For products targeting domestic users, real device testing is an indispensable step. Minimum requirements:
- Huawei (Hongmeng OS with WeChat WebView)
- Xiaomi (MIUI system)
- iPhone (WeChat WebView vs. Safari)
You can use BrowserStack's real device cloud testing service or collaborate with a well-equipped testing team.
Chapter 8: Honest Analysis — What Can Be Built, What Will Be Fragile, What Requires a Developer
What Domestic Developers Can Reliably Build Without Writing Code
The following product types have been successfully built by many Chinese developers using AI tools, with high reliability:
✅ WeChat Mini Programs (basic functions) Display, form, appointment Mini Programs — TRAE CN's understanding and generation quality is sufficient.
✅ Internal management systems (operation backends) Data display, user management, order management backends — both Tongyi Lingma IDE and CodeBuddy have good support.
✅ Content websites / landing pages Brand websites, event landing pages, blog systems — AI tools generate quickly and with stable quality.
✅ Basic SaaS tools (single core function) SaaS focused on one core function (e.g., "document to PDF", "resume generator") — can be completed without deep technical knowledge.
Can Be Built, But Needs Careful Review
⚠️ Payment integration AI can generate the basic framework, but payment security (signature verification, idempotency handling, reconciliation logic) is high-risk. A technical person must review every line of payment-related code.
⚠️ Real-name verification Three-factor authentication (name + ID + face recognition) via Alibaba Real-Person Verification or Tencent Cloud Face Recognition. AI can generate the calling code, but the business logic (when verification is mandatory, how to handle failure) needs human design.
⚠️ High-concurrency flash sale / rush purchase functionality Combination of Redis + distributed lock + message queue. AI can propose a solution, but load testing and edge conditions must be manually validated.
⚠️ Products involving minors Domestic regulations on minor protection are strict (anti-addiction system integration requirements). This is not just a technical issue but involves legal compliance.
Requires Experienced Engineers
❌ Financial products (loans, wealth management, payments) Licensing requirements + anti-money laundering compliance + fund custody — complex systems involving multiple regulatory bodies. AI tools cannot handle compliance decisions.
❌ Healthcare products (internet hospitals, online consultations) Qualification requirements for internet hospitals / drug sales and data localization requirements go far beyond technical issues.
❌ Live streaming / short video e-commerce systems Content moderation (image recognition, NLP filtering) + complex commission sharing systems + live streaming tech stack — each alone is a specialized engineering domain.
❌ Large enterprise private deployment Security auditing, data migration, integration with existing legacy systems — requires solution architects with private deployment experience.
What Chinese Developers Think Needs an Engineer But Actually Doesn't
✅ SMS verification code login: Alibaba Cloud / Tencent Cloud SMS + AI-generated login logic — highly reliable.
✅ Basic WeChat Pay integration: As long as you use the official SDK + thorough testing, AI-generated code is usable for simple scenarios.
✅ OSS / COS file upload: AI's understanding of such standardized API calls is very accurate.
✅ Basic data dashboard: Echarts + SQL queries + Next.js API — AI can directly generate usable code.
Chapter 9: Getting 80-90% of Premium Tool Performance for ¥0-40 per Month
China market's unique ultra-low-cost solutions.
Why the Cost Ceiling Is Lower for Chinese Developers
Compared to international markets, Chinese developers have unique cost advantages:
- Many domestic AI-native IDEs are free: TRAE CN basic edition is free, with functionality already far exceeding Cursor's free edition.
- Coding Plan breaks token billing deadlock: The first-month ¥7.9 Coding Plan makes agent-level usage cheap.
- DeepSeek API costs crush international competitors: At equivalent performance levels, the price is 1/10 that of Claude Sonnet.
Zero-Cost Solution (¥0/month)
Combo: TRAE CN free edition + CodeGeeX plugin (VS Code)
- TRAE CN free: Doubao models free, basic AI features complete
- CodeGeeX: 2,000 free requests per day, supports VS Code and JetBrains
- Suitable for: code completion, chat Q&A, simple code generation
Limitations: Doubao models have limited complex task capability; CodeGeeX completion quality is weaker than TRAE's Cue feature.
Suitable for: Students, entry-level developers, light usage scenarios.
Ultra-Low-Cost Solution (¥7.9 first month, then ¥39.9/month)
Combo: OpenCode terminal tool + Alibaba Cloud Bailian Coding Plan
Setup steps:
# Step 1: Install OpenCode
# macOS / Linux
curl -fsSL https://opencode.ai/install.sh | sh
# Windows (PowerShell)
irm https://opencode.ai/install.ps1 | iex
# Step 2: Configure Alibaba Cloud Bailian API key
opencode config set provider bailian
opencode config set api_key sk-xxxx # get from Bailian console
# Step 3: Choose model (Qwen3-Max recommended)
opencode config set model qwen-plus-coding
Actual value of Bailian Coding Plan: ¥39.9/month includes unlimited calls (within quota) to 8 mainstream models, covering Qwen3-Max, MiniMax M2.5, GLM-5, Kimi K2.5, etc. For moderate users, this is the most cost-effective solution currently.
Achieves 80% of premium tools:
- ✅ Complex single-file and multi-file code generation
- ✅ Bug diagnosis and fixing
- ✅ Code refactoring and optimization suggestions
- ✅ Terminal command generation and scripting
What it cannot do:
- ❌ TRAE's SOLO Mode (full-process automation)
- ❌ Cursor's smooth Tab completion experience
- ❌ Graphical code diff interface
Low-Cost High-Performance Solutions (¥40-80/month)
Option 1: TRAE CN Individual Edition (¥49)
The best value "all-in-one" choice. ¥49/month includes advanced model access (DeepSeek, Kimi, etc.), priority queue, and full SOLO Mode experience. For most full-time developers, this is the most suitable starting point.
Option 2: OpenCode + DeepSeek V4 Pro API direct
DeepSeek V4 Pro currently enjoys 75% discount, input price ~$0.075/M tokens, output ~$0.125/M tokens. During the discount period, this is an extremely cost-effective choice, with code quality close to Claude Sonnet level but at much lower cost.
Average monthly cost: ¥50-100 (moderate use)
When "Saving Money" Starts to Cost You
Signals to upgrade to a premium solution:
- You stop for more than 30 minutes each day because you hit the Coding Plan quota limit.
- You find yourself planning "which model saves more money" instead of focusing on business problems.
- You need to work on large codebases with hundreds of thousands of lines (requiring larger context windows).
- Your team has more than 3 people using AI coding tools simultaneously (individual plans don't support multi-user collaboration management).
Chapter 10: Outlook for China's AI Coding Market in the Next 18 Months
Five Structural Changes Underway
Change 1: Domestic models are entering the truly "good enough" range
The statement "China has caught up to the West in coding LLMs" needs a footnote: perhaps in synthetic reasoning benchmarks; in delivering a complete application where every part runs, one model (Kimi K2.6) has caught up, and the rest still lag a generation behind.
This is an honest assessment. But the fact that "one model caught up" is itself significant — it shows that Chinese teams, after increasing investment in models, have been able to break through previous bottlenecks.
Prediction: 18 months from now, 3-5 Chinese models will reach the "fully usable" standard in real-world application building tests. At that point, Chinese developers' cost advantage will become even more pronounced.
Change 2: The Coding Plan subscription model will become a global standard
The Chinese market invented the Coding Plan subscription logic, solving the bill shock problem in agent coding scenarios. This model is already influencing international markets — Anthropic's Max subscription, Cursor's enterprise plans — all evolving in a similar direction.
Prediction: 18 months from now, pay-as-you-go API billing will be used more for development and testing, while production-level AI coding will primarily rely on subscription packages.
Change 3: Big tech's free strategy is unsustainable; the paid ecosystem will gradually mature
ByteDance's free strategy led the way, and other Chinese companies had to follow. As of now, Baidu Wenxin Kuaima Comate Individual Standard is free; Tencent CodeBuddy is time-limited free for individual developers; Alibaba Cloud Tongyi Lingma Individual is free, and Individual Pro is also time-limited free; Huawei CodeArts Individual Basic is free.
The words "time-limited free" mean change is brewing.
Prediction: 12-18 months from now, most "time-limited free" products will shift to paid, but prices will remain in the ¥29-99/month range (significantly lower than international competitors), forming a uniquely Chinese pricing ecosystem.
Change 4: The enterprise market is the real battleground
Profit margins in the C-end price war are extremely thin. The true ROI for big tech comes from enterprise clients. Among enterprise customers, Alibaba's Tongyi Lingma is most commonly seen — not only because the product is competitive, but also because Alibaba's Qwen LLMs are well-developed and its B2B services are attentive.
Prediction: The enterprise AI coding market will enter a truly competitive phase in the next 18 months. MLPS certification, private deployment, and code security auditing will become decisive competitive factors.
Change 5: Agent mode will shift from novelty to standard
After TRAE China edition's SOLO Mode launched, 3 out of every 10 TRAE developers adopted SOLO.
30% penetration, just after a new feature launch. This indicates that acceptance of agent mode is far exceeding expectations.
Prediction: 18 months from now, over 60% of domestic AI coding tool users will use agent mode at least once a week. "Directing AI to complete tasks" will become the norm.
Response Strategies for Different Roles
Non-technical founders:
Start validating your product idea with TRAE CN SOLO Mode right now. Don't wait until you "learn to code." First use AI to build a demonstrable version, then use that version to find a technical co-founder or angel investment. A clickable demo is worth far more than a slide deck.
Junior/mid-level engineers:
The opportunity and challenge for you: AI tools can bring your output close to that of senior engineers, but at the same time, senior engineers using AI tools also see their output significantly increase.
Core advice: Don't just use AI to write code — use AI to understand system design, learn architectural patterns, and supplement areas of technology you're unfamiliar with. Use AI tools as learning accelerators, not just code generators.
Senior engineers / architects:
Your core value is shifting — from "writing excellent code" to "designing system constraints that AI can correctly execute." This is a brand new skill: how to decompose complex engineering problems into subtasks that AI can reliably complete, how to design .cursorrules and CLAUDE.md so AI follows architectural decisions long-term, and how to define "what constitutes correct code."
Enterprise technology leaders:
Your most important decision now is not "which AI coding tool to use," but "how to design engineering processes for AI-assisted development" — including how code review adapts to AI-generated code, how to shift security and compliance left, and how to keep technical debt manageable while AI accelerates development.
Honest Answer to "Can Domestic AI Tools Replace Cursor?"
Reality as of May 2026: Not completely, but the gap is closing fast.
TRAE CN can already reliably replace Cursor in these scenarios:
- Projects with Chinese code comments and documentation
- WeChat/Alipay ecosystem development
- Medium-complexity full-stack web applications
Scenarios where TRAE CN still lags behind Cursor:
- Global understanding of ultra-large codebases (>500k lines)
- Fluency and accuracy of Tab completion
- Ecosystem maturity (plugin count, community resources)
18 months from now: If Doubao models continue to iterate at current speed, and with TRAE's continued investment in agent capabilities, there is a high probability that TRAE CN's comprehensive experience for domestic development scenarios will reach Cursor's level by the end of 2027.
Conclusion: Start Now, Not Later
The landscape of AI coding tools in China is more favorable to developers than any other market in the world: tools are free, models are accessible domestically, and prices are one-tenth of international markets.
This is a rare window in history — the boundaries of technical capability are expanding rapidly, but the first-mover advantage has not yet been fully seized.
Based on your current situation, choose an entry point:
- If you have never written code: Register for TRAE CN, open SOLO Mode, describe the smallest tool you want in Chinese, and watch how it gets built.
- If you are a junior developer: Install TRAE CN or Tongyi Lingma plugin, use it to write a feature in your current project. Compare: which is faster, which has better quality — you writing vs. AI-assisted.
- If you are a senior developer: Set up OpenCode + Alibaba Cloud Bailian Coding Plan, use ¥7.9 for a month of agent-level experience. Observe how far it can go in your codebase, then consider its place in your workflow.
Appendix: Quick Reference Card for Chinese Developers
| My Situation | Recommended Tools | Monthly Cost | First Step |
|---|---|---|---|
| Non-technical, validating ideas | TRAE CN free (SOLO Mode) | ¥0 | Download TRAE CN, input product description |
| Junior developer, daily assistance | TRAE CN free + Tongyi Lingma plugin | ¥0 | Install both, feel the difference |
| Full-time developer, primary IDE | TRAE CN Individual | ¥49 | Upgrade to Individual, enable DeepSeek models |
| High performance + low cost | OpenCode + Alibaba Cloud Bailian Coding Plan | ¥7.9 first month | Install OpenCode, buy Bailian Lite Plan |
| Enterprise team, no compliance requirements | TRAE CN Enterprise | Contact sales | Apply for Enterprise trial |
| Enterprise team, MLPS requirement | Tongyi Lingma Enterprise | from ¥79/person/month | Contact Alibaba Cloud enterprise service team |
| Strong technical, extreme control | OpenCode + Yueli-KGM-computing or Ollama/VLLM (Qwen3-32B) | ¥0 | Install Ollama, pull qwen3:32b model |
| Domestic developer using overseas tools | Cursor Pro + Volcano Ark Coding Plan | $20 + ¥49 | Use Volcano Ark API instead of Anthropic API |
This guide is based on tool status and market data as of May 2026. The AI tools space changes extremely fast — it is recommended to follow the official changelogs of TRAE, Tongyi Lingma, CodeBuddy, etc. Domestic Coding Plan prices fluctuate frequently, refer to each platform's official website for real-time prices.