Stack IA du Développeur en Chine — Coder Derrière le GFW

What's in this pack

This is the rig for a working engineer in mainland China who hits the same three walls every week: Cursor and Codex sign-up requires a foreign phone, OpenAI / Anthropic APIs are intermittently blocked at the network layer, and corporate firewalls + WeChat-Pay-only billing make foreign SaaS a non-starter. The pack is not a workaround for those tools — it's the native-China alternative built on Chinese model labs (DeepSeek, Alibaba Qwen, Moonshot Kimi) plus self-hostable gateways and IDE plugins.

Every pick here meets three criteria: RMB-payable or self-hostable, reachable from a typical mainland network without a VPN, and good enough to actually use as a daily coding assistant — not just "works". If you have Cursor working and don't want to switch, this isn't your pack. If you're on a company-issued laptop with no foreign network access, this is the only pack that ships.

Install in this order

1. DeepSeek TUI (id: 3142) — start here because DeepSeek-V3 / R1 is the highest-quality Chinese-trained model with an RMB-billable API, and the TUI is a one-binary terminal agent that proves the API key works before you wire anything else up. npm i -g deepseek-tui, paste your key from platform.deepseek.com, get answers in 30 seconds.
2. Qwen Code (id: 3022) — Alibaba's official terminal coding agent for Qwen models. Install second because Qwen-Max / Qwen3-Coder are the strongest alternative when DeepSeek capacity is throttled, and you want a fallback model wired up before you write any production scripts. npm i -g @qwen-code/qwen-code, /auth with your DashScope (bailian.console.aliyun.com) key.
3. oh-my-kimi (id: 3643) — Moonshot's Kimi K2 has the longest reliable context window (1M tokens) among China-hosted models, and oh-my-kimi gates the agent loop with evidence checks. Install third to round out the model layer: DeepSeek for daily code, Qwen as backup, Kimi for whole-codebase questions.
4. Cherry Studio Custom Models (id: 2821) — open-source desktop chat client (Mac/Win/Linux) that adds any OpenAI-compatible endpoint as a BYOK provider. Install fourth because once you have three model providers, you need one chat UI that can flip between them mid-conversation. Cherry Studio is the de facto standard in Chinese dev communities — 30K+ stars, MIT license, no telemetry.
5. One API (id: 3821) — self-hosted LLM API gateway (Docker, Go). Install fifth because by now you have multiple keys (DeepSeek + Qwen + Kimi) and your team needs a single OpenAI-compatible URL that round-robins across them. One API is the gateway most Chinese teams reach for first (33K+ stars, Chinese-language admin UI).
6. LiteLLM Proxy (id: 2789) — the international alternative gateway, Python-based, 100+ provider support. Install as an alternative to One API only if you also need to proxy international providers (OpenRouter, Together, Anthropic via reverse-proxy) and want richer cost tracking. Don't run both.
7. Continue (id: 613) — open-source IDE assistant (VSCode + JetBrains) that points at any OpenAI-compatible endpoint — which means it points at your One API gateway, which means your IDE autocomplete and chat are now powered by DeepSeek-Coder for ¥0.50 per million tokens.
8. Tabby (id: 216) — self-hosted GitHub Copilot alternative. Install only if your company forbids external API calls (banks, gov, internal R&D). Tabby runs entirely on your hardware, supports DeepSeek-Coder / Qwen-Coder weights, and gives FIM autocomplete in VSCode + JetBrains.
9. Ollama (id: 162) — local model runtime. Install if you have a Mac with 32GB+ unified memory or a workstation with a 24GB+ GPU. Run ollama pull qwen3:14b or deepseek-r1:7b and Cherry Studio / Continue can talk to localhost:11434 instead of any API.
10. Reasonix (id: 3604) — DeepSeek-native coding agent CLI optimized for prompt caching (claimed 99%+ cache hit). Install last as the power-user upgrade once you've used DeepSeek TUI for a week and want lower latency + lower bills on a real repo.

How they fit together

  Domestic model APIs (all RMB-billable, no VPN needed)
  ┌─────────────────┬─────────────────┬─────────────────┐
  │ DeepSeek-V3/R1  │  Qwen-Max/Coder │  Kimi K2 (1M ctx)│
  │ (deepseek.com)  │  (bailian.aliyun)│ (platform.moonshot)│
  └────────┬────────┴────────┬────────┴────────┬─────────┘
           │                 │                 │
           ▼                 ▼                 ▼
           One API  /  LiteLLM Proxy  (Docker, your VPS)
           (unified OpenAI-compat URL + key rotation + cost log)
                              │
        ┌─────────────────────┼─────────────────────┐
        ▼                     ▼                     ▼
  DeepSeek TUI /         Cherry Studio        Continue (VSCode/JB)
  Qwen Code /           (desktop chat,         (in-IDE chat
  oh-my-kimi /          BYOK any provider)     + completions)
  Reasonix
  (terminal agents)

  ── Air-gapped fallback (no external network at all) ──
          Tabby (self-hosted Copilot, runs DeepSeek-Coder weights)
          Ollama (local Qwen / DeepSeek-Coder / Yi / GLM weights)

The inflection point is One API in front of three domestic providers. Before that, you have three API keys scattered across .env files and every tool has to know about each provider. After that, every tool points at one URL, you swap models by renaming the model field, and your billing team sees one consolidated cost report. Don't skip the gateway — without it, multi-model workflows fall apart by week two.

Tradeoffs you'll hit

• DeepSeek vs Qwen vs Kimi — DeepSeek wins on code quality + price (¥0.50-2/M tokens). Qwen wins on Chinese-language documentation, multi-modal, and Alibaba Cloud enterprise compliance. Kimi wins on context length (1M, actually usable). Most teams run DeepSeek default, Qwen as Aliyun-compliance fallback, Kimi for codebase-wide Q&A.
• One API vs LiteLLM — One API is Chinese-developer-first (UI in Chinese, supports WeChat Pay billing plugins, lighter footprint). LiteLLM is international-first (better cost tracking, more provider integrations, Python ecosystem). If your only providers are Chinese, One API is the obvious pick.
• Continue vs DeepSeek TUI / Qwen Code — Continue lives in your IDE (autocomplete + chat panel). The CLI agents live in terminal and can execute multi-step refactors. They complement each other: Continue for typing, CLI agents for "refactor this whole module". Most engineers run both.
• Cloud API vs Tabby vs Ollama — Cloud API is cheapest and highest quality but requires external network. Tabby is mid-quality (model is local) and needs a GPU. Ollama is your laptop; quality drops sharply below the 14B parameter mark. Only go local if compliance demands it or you genuinely have a 4090 / M-series Pro burning idle cycles.
• Cherry Studio vs the model labs' web chat — DeepSeek, Qwen, and Kimi all have free web chat. Cherry Studio is worth installing the moment you want to switch models mid-conversation, save chat history locally, or pipe in your own RAG knowledge base. If you only use one model from a browser, you don't need it.

Common pitfalls

• "DeepSeek API is down" — DeepSeek has had multi-hour capacity issues during peak periods. This is exactly why One API + a Qwen fallback exists. Configure the gateway with priority: 1 deepseek, priority: 2 qwen and most outages become invisible.
• JetBrains Continue plugin marketplace can be slow from CN — if the plugin fails to install from inside an IDE, download the .zip from plugins.jetbrains.com (often reachable) and install from disk. VSCode marketplace is generally fine via Microsoft's CN mirrors.
• Bailian (Aliyun) auth flow assumes a Chinese phone — Qwen API via bailian.console.aliyun.com requires Aliyun account + Alipay real-name verification. If your team is on a foreign Aliyun account, use Qwen via DashScope international (dashscope-intl.aliyuncs.com) — different endpoint, same model, different billing entity.
• Cherry Studio + corporate proxy — Cherry Studio respects system proxy by default, which breaks DeepSeek calls if your proxy whitelist only includes foreign domains. Either add api.deepseek.com to the proxy bypass list, or configure Cherry Studio's per-provider proxy override.
• Tabby + GPU memory — DeepSeek-Coder-V2-Lite-16B needs ~24GB VRAM at fp16, ~12GB at 4-bit quant. A single 3090 / 4090 is enough; a 4060Ti 16GB is not. Right-size before promising your team Copilot replacement.
• One API on a public IP without auth = key theft — One API admin UI defaults to root/123456. Change it the first hour, put it behind nginx + basic auth or a VPN. Several Chinese dev forums have screenshots of stolen DeepSeek keys racking up ¥10K bills overnight.
• WeChat Pay / Alipay billing topup minimums — DeepSeek and Moonshot have ¥10-50 minimum top-ups; not a problem for individuals, but for company finance teams that need fapiao (发票), request the developer/enterprise account from day one — switching mid-stream wastes a week.