China AI API Alternatives — OpenAI / Claude Replacements You Can Pay in RMB

What this pack actually solves

A team in mainland China shipping an AI feature hits the same three walls. (1) The OpenAI / Anthropic billing flow needs a foreign card and a foreign-IP signup — practical for a side project, painful at company scale. (2) Outbound API calls to api.openai.com / api.anthropic.com are unreliable from inside China without a proxy, and a corporate proxy raises its own compliance questions. (3) Finance wants a fapiao (增值税发票) for the spend, which foreign SaaS does not issue.

This pack picks ten assets that close all three gaps without giving up GPT-4o-class quality. Three tiers: SOTA frontier (DeepSeek-V3, R1, Coder), Chinese-vendor alternates (Qwen, ChatGLM, Kimi, MiniMax), and gateways that hide vendor choice behind one OpenAI-compatible endpoint (One API, LiteLLM Proxy, Cherry Studio). End state: your code stays unchanged — base_url and model are the only diffs — and finance gets a fapiao every month.

Install in this order — free / cheap first, then quality, then routing

1. DeepSeek-V3 — Open-Weight 671B MoE with GPT-4o Quality (id 2832). Start here. The hosted API is $0.27 / 1M input tokens — roughly 10× cheaper than GPT-4o — billed in RMB at platform.deepseek.com. OpenAI-compatible: switch base_url to https://api.deepseek.com/v1, set model="deepseek-chat", ship. Weights are MIT-licensed if you ever want to bring it in-house.
2. Qwen Code — Terminal Coding Agent for Qwen Models (id 3022). Alibaba Cloud DashScope hosts Qwen2.5 / Qwen3 with OpenAI-compatible endpoints and RMB billing. Qwen Code is the CLI wrapper most domestic devs use to wire Qwen into a coding loop without writing the SDK glue themselves. Pair the API with this CLI on day one.
3. ChatGLM — Open Bilingual Chat Model by Tsinghua KEG (id 2264). The classic Zhipu / Tsinghua bilingual line. The hosted GLM-4 series on bigmodel.cn ships in RMB, supports fapiao, and remains the strongest non-DeepSeek choice when the workload is bilingual customer-facing — translation, support, content moderation in mixed zh/en.
4. oh-my-kimi — Evidence-Gated Agent Runtime for Kimi (id 3643). Moonshot's Kimi line is the long-context champion in the domestic market (200K+ tokens, document-grade comprehension). Wire it for the "feed in the whole contract / RFP / manual" workload that GPT-4o handles uneconomically. oh-my-kimi adds the agent-runtime layer so the model is not just a chat box.
5. MiniMax-MCP — Official MiniMax MCP Server (id 3932). MiniMax's strength is voice and multi-modal (TTS, voice cloning, video-to-text). The official MCP server gives any MCP-aware client (Claude Desktop, Cursor, Cline) a tool surface for those models. Install this once you need anything beyond text.
6. DeepSeek-R1 — Open-Weight Reasoning Model Rivaling OpenAI o1 (id 2833). The reasoning sibling of V3. Same API surface, different model alias (deepseek-reasoner). Route hard reasoning calls here, route everything else to V3 — the gateway in step 8 makes that one config line.
7. DeepSeek Coder — Code-Specialized Model for Local Inference (id 2834). The smaller, code-specialized branch — runs locally on a workstation GPU and removes the network hop entirely for code completion or in-IDE refactor. Belongs on a developer's laptop, not in your production gateway.
8. One API — Unified LLM API Gateway (Docker) (id 3821). The single most important install in this pack. One API is the open-source, China-friendly OpenAI-compatible gateway — it speaks the OpenAI request/response format and routes calls to DeepSeek, Qwen, ChatGLM, Kimi, MiniMax, Azure, Anthropic, and dozens more. Run it in Docker on your own VPC, point your code at it, and switching vendors becomes a config row, not a code change. RMB billing for downstream providers is preserved.
9. LiteLLM Proxy — Unified Gateway for 100+ LLM APIs (id 2789). The Western-built equivalent. Use it instead of One API if your team already speaks Python, wants per-key cost tracking out of the box, or needs the more mature failover and rate-limit logic. Same shape — one OpenAI-compatible endpoint, many providers behind it.
10. Cherry Studio Custom Models — BYOK Any LLM Provider (id 2821). The desktop client. Before you wire any of the above into production, paste your keys into Cherry Studio and sanity-check that the model actually answers in the way your app expects. It is the cheapest way to compare DeepSeek vs Qwen vs Kimi on real prompts before you commit a routing decision.

How they fit together

          [ Frontier tier ]              [ Vendor-alternate tier ]      [ Local fallback ]
          DeepSeek-V3 (general)          Qwen (bilingual / Alibaba)     DeepSeek Coder (laptop)
          DeepSeek-R1 (reasoning)        ChatGLM (Zhipu / Tsinghua)
                  │                      Kimi  (long context, Moonshot)
                  │                      MiniMax (voice / multi-modal)
                  │                              │
                  └────────── One API ───────────┤
                              (Docker, OpenAI-compatible)
                                       │
                              LiteLLM Proxy (alt routing layer)
                                       │
                                Your app code
                                       │
                              Cherry Studio (dev sanity check)

The spine is DeepSeek-V3 + One API + LiteLLM Proxy + Cherry Studio. That quartet covers the general workload, hides vendor choice behind one endpoint, gives you a routing layer with cost tracking, and gives a desktop client for manual eyeballing. Qwen / ChatGLM / Kimi / MiniMax / DeepSeek-R1 / Coder are alternate backends you add behind the gateway as the workload demands a different specialty.

Tradeoffs to know before you commit

• Context length. Kimi leads (200K+), Qwen and DeepSeek sit in the 64K-128K band for most plans, GLM-4 around 128K. If the workload is "summarize this 80-page PDF," Kimi is the only domestic answer that does not require chunking.
• English quality. All five families speak English. DeepSeek and Qwen are closest to GPT-4o on English benchmarks; GLM and Kimi are slightly behind on creative English but ahead on Chinese. Test on your prompts — benchmark averages do not predict your workload.
• Pricing units. Most providers price per million tokens in RMB on their dashboard and in USD on the public API page. The numbers are not always one-to-one — sometimes the RMB rate is cheaper, sometimes not. Always log usage to your own ledger; do not trust the dashboard alone for cost reconciliation.
• Rate limits and concurrency. Frontier models on a new account often start at 60 RPM / 1 concurrent request. Production workloads need a tier-up application — sometimes a phone call. Budget a week of lead time before launching anything that needs > 10 RPS.
• Reasoning model latency. R1, GLM-4-Reasoning, Qwen-QwQ trade speed for quality. A reasoning call can take 30-90 seconds — wire a streaming response or a job queue, do not let it block a synchronous request.

Common pitfalls

• Hardcoding the vendor SDK. If you import from openai import OpenAI everywhere and only switch base_url, you stay portable. If you import dashscope or zhipuai directly, you are now coupled to one vendor and cannot route through a gateway. Use the OpenAI SDK against the OpenAI-compatible endpoint every vendor publishes.
• Treating ICP / 备案 like an API concern. ICP filing is for the public-facing site hosted in China, not for the API call itself. You do not need a 备案 to call DeepSeek's API from a server outside China. You do need a 备案 if your domain is *.cn and serves users in China. Keep the two questions separate.
• Forgetting that token counting differs. OpenAI tokenization (tiktoken / cl100k) does not match DeepSeek / Qwen / GLM tokenization. A 1000-character Chinese prompt that is 350 OpenAI tokens may be 280 DeepSeek tokens. Cost projections built on tiktoken will be off by 10-30% — always count with the vendor's own tokenizer for the budget that matters.
• Running tests against the production endpoint with the prod key. Every vendor in this pack has a sandbox / test key. Use it. Otherwise the first time your CI loops gets you a $200 surprise from a Qwen-Max test.
• Skipping the fapiao setup. The major vendors (DeepSeek, Alibaba Cloud, Zhipu, Moonshot, MiniMax) all support fapiao but each has a different form. Get finance to file the request the same week you sign up — the first month's invoice cycle takes 30 days, and they will not back-issue.
• Assuming deepseek-chat is always V3. Vendor aliases change. deepseek-chat points to the current chat flagship; today that is V3, tomorrow it could be V3.5 / V4. If your eval depends on a specific weight, pin the explicit version string, not the alias.