This is a field guide, not a sales pitch. I wrote it because every "how to use Chinese LLMs outside China" post I found was either outdated (2024-era), a thin affiliate link, or assumed you already know what a "peak/off-peak pricing window" is. If you're an overseas developer who just wants DeepSeek/GLM/Qwen to work with a normal credit card, this is the post I wish I'd had.
The problem you actually hit
You've heard DeepSeek-V4, GLM-5, and Qwen3 are excellent and cheap. You try to use them. Then:
-
The official site asks for a
+86phone number.platform.deepseek.comonboarding is gated for mainland China. Same story for several first-party endpoints. - It wants RMB and a local payment method. Even when you get in, topping up usually needs a Chinese bank card or Alipay/WeChat Pay.
- The open-web tutorials are stale. Most "use DeepSeek from abroad" articles were written before the 2025–2026 pricing and access changes. They point you at endpoints that now behave differently.
- The reseller market is a maze. OpenRouter adds a markup. SiliconFlow shows peak/off-peak windows (e.g. DeepSeek-V4-Flash at ¥3.00/¥9.00 in/out at peak, ¥1.50/¥4.50 off-peak) that make your cost prediction impossible. You end up not trusting any of it.
None of these are about model quality. They're about access friction. That's the real product gap — not "who is 4% cheaper."
Three roads, honestly compared
There are exactly three ways to get these models from outside China. Here's what each actually costs you in effort.
Road A — Official international endpoints
Good news most people miss: several Chinese model providers already run international endpoints that do NOT require a +86 number. They take an email + an international card.
- DeepSeek:
api.deepseek.com(international) - Zhipu GLM:
api.z.ai - Alibaba Qwen:
dashscope-intl - Moonshot, ByteDance: international consoles exist
Pros: you buy straight from the source, no middleman. Cons: you must register each provider separately, manage separate API keys, separate billing, and learn each one's quirks. If you use three models from three vendors, that's three dashboards, three invoices, three rate-limit policies. For a solo dev shipping a product, that overhead is real.
Road B — Self-host the open weights
DeepSeek, Qwen, and several GLM variants are open-weight. You can run them on a cloud GPU with vLLM or sglang.
pip install vllm
vllm serve Qwen/Qwen3-235B-A22B --tensor-parallel-size 4
Pros: total control, no per-token markup, data stays on your infra. Cons: you need GPU budget, you run the serving stack (autoscaling, cold starts, quantization choices), and you pay for idle time. This is the right call at large scale or for sensitive data — not for "I want to prototype this week."
Road C — A unified gateway / aggregator
One API key, one OpenAI-compatible endpoint, one invoice, routing to many models.
Examples: OpenRouter, Novita, and (full disclosure) TideLink, which is the one I run.
Pros: one integration, one bill, one support contact. Cons: you're trusting a middleman, and — fairly — you pay for that convenience. The legitimate question is how much and how transparently.
The question everyone actually has: "am I being ripped off?"
Fair. Here's the honest framing the reseller market avoids.
A gateway's price = upstream cost × a margin. The upstream cost for these models genuinely fluctuates (provider price changes, peak/off-peak windows, wholesale tiers). So a transparent gateway should be able to show you its margin, not hide behind "cheapest ever!!" banners.
At TideLink we do exactly that: we publish upstream cost × 1.30 as the price, no mystery tier. You can verify which model actually served your request:
from openai import OpenAI
client = OpenAI(
base_url="https://tidelink.xyz/v1",
api_key="YOUR_KEY", # get a free tier key, no card required for GLM
)
resp = client.chat.completions.create(
model="glm-4.7-flash", # free tier, real model
messages=[{"role": "user", "content": "ping"}],
)
print(resp.model) # <- the ACTUAL model id served
print(resp.usage) # <- real token counts
That resp.model line is the whole trust story: a gateway that hides which model it served is a gateway you shouldn't trust. Ask any provider for it; if they can't show it, that's your answer.
A concrete, working setup for an overseas solo dev
If I were starting fresh today, here's what I'd do — and it costs $0 to try:
-
Start on the free GLM tier.
glm-4.7-flashis genuinely free via TideLink's gateway — no card, no+86. Good enough for prototyping, RAG, classification, and a lot of production workloads. -
Verify the model is real with the
resp.modelsnippet above. (It is. We don't swap it.) -
When you need DeepSeek/flagship quality, you pay
upstream × 1.30. You see the number before you call. - One invoice, in USD, at month-end — useful if you're billing a client or doing accounting, which most of the first-party Chinese endpoints won't give you in a form your accountant accepts.
That's it. No Chinese phone number. No RMB. No KYC wall. No three dashboards.
What I'm NOT claiming
- I'm not claiming we're the cheapest. On flagship DeepSeek we are not — and anyone who tells you they're 6× cheaper is either on a wholesale tier you can't get, or not being straight about which model they serve.
- I'm not claiming we replace self-hosting at scale. If you're doing millions of tokens/day, Road B will beat any gateway. That's math, not pride.
What we are is the lowest-friction honest on-ramp: free GLM entry, transparent cost × 1.30, a real resp.model, and a USD invoice. If that's the gap you hit, the road is here.
Written by the team behind TideLink. If something in this guide is wrong or outdated, open an issue or email support@tidelink.xyz — we'd rather fix it than let a stale tutorial mislead someone. Attribution tag for our own analytics: ?cid=devto.
Get a free TideLink API key — call GLM, Qwen, DeepSeek and more through one OpenAI-compatible endpoint, no card required for the free GLM tier: https://tidelink.xyz/dashboard.html?cid=devto
