Three APIs that claim to require overseas networks can be used directly in China

Let's make it clear what this article tested: my own model gateway is equipped with 13 free large model entrances, and this time I tested these 13, not a census of all free APIs on the market. There are far more free suppliers that can be directly connected in China. Domestic platforms such as silicon-based mobility, Bailian, and Qianfan do not have network problems and are not included in this time.

There is another thing to measure. When it comes to Groq, Cerebras, NVIDIA, and OpenRouter, most strategies will add the sentence "requires an overseas network environment." I classified it this way in the previous articles myself, because the server has always been equipped with overseas exports, and I have never verified whether this sentence is correct.

Yesterday, I happened to have a clean domestic machine in my hand: the original production server was in the domestic computer room, and the export configuration had been completely cleared. So I tested these platforms one by one and made them purely connected.

The results don't match popular wisdom.

!

Among the 13 entrances ## , 4 domestic direct connections can be adjusted.

It's not that you can ping, but you can send a real reasoning request with a key, get a 200 and a complete reply:

platform	time-consuming	Does it conform to popular wisdom?
Smart Spectrum GLM-4-Flash	1.11s	It is consistent, it is a domestic platform
Cloudflare Workers AI	1.85s	not meet
OpenRouter	3.22s	not meet
NVIDIA NIM	22.08s	not meet

In most strategies, the latter three companies are classified as "overseas platforms that require overseas network environment", and they are actually available directly. This is the only discovery worth mentioning this time: it is not that "only these four entrances in China can be used", but that "among the few companies labeled as needing overseas networks, three of them actually don't need it."

NVIDIA's 22 seconds have to say separately: it worked, but it was ridiculously slow. For the same model, I tested 4 seconds on a server that went overseas, and 22 seconds for a direct connection within China, which is more than five times the difference. So "usable" and "easy to use" are not the same thing here. It is more suitable for running offline tasks and not suitable for placing them in places where people are waiting.

batch is actually blocked these four

Cerebras and Groq both return 403, but the nature is different, depending on the return.

Cerebras 'response read:

HTTP/2 403
server: cloudflare
error code: 1009

1009 is the error code Cloudflare uses to deny access by region. This is not an authentication failure without a key, nor is it a traffic restriction, but the service party has made access restrictions according to IP attribution.

Groq is 403 Forbidden, which is also returned by Cloudflare. cf-ray shows that it is going to the Singapore node.

Mistral and Gemini are even more straightforward. The connection times out directly and the TCP layer becomes inaccessible.

Therefore, these companies do need an overseas network environment, which is consistent with popular saying.

still has 4, it is not the network stuck

This category is the most easily misjudged.

Kimi returned 429, ModelScope returned insufficient balance, Volcano returned 401, SambaNova's model list can be pulled normally, but the reasoning requires binding the payment method first.

What they have in common is that the network layer is completely smooth, and what is stuck is the account. The balance is gone, the amount is in arrears, and the card needs to be tied.

This distinction is very important, because when reporting an error, people's first reaction is often "Is the network down?" So they go to trouble the network environment, but the actual stuck point is at the end of the account. The judgment method is also simple: first request the /v1/models interface of this house. If you can get the model list, it means that there is no problem with the network. Go to the account background to check the rest.

!

By the way, registration threshold

Since we talk about whether we can use it or not, we will also talk about the hurdle of registration. This part was obtained by checking the public documents and registration pages of each family. I did not register and verify it one by one. Let me explain the source.

According to public information, Groq, Cerebras, NVIDIA NIM, Google Gemini, and OpenRouter's free layersdo not need to bind a credit card, email or Google account registration to get an API Key. Mistral's Codestral free credit also does not require binding cards.

To be specific: Cerebras free layer is 1 million tokens per day at 30 RPM;NVIDIA registers for 1000 inference quotas and can apply for expansion to 5000 at a rate of 40 RPM;Mistral's Codestral is 1 billion tokens per month but only 2 RPM. These figures come from public explanations from various families, and I have not verified them one by one.

Therefore, for domestic users, the threshold is actually not about money or card binding, but about two things: whether the registration page can be opened, and whether the call will be rejected due to the region after registration. The table above has already answered these two matters.

, so how to match it is the most convenient

Looking at accessibility and registration thresholds together, if you are in China and don't want to mess with the network environment, in addition to domestic platforms that are already fine, these overseas entrances can also be directly used:

Smart Spectrum GLM-4-Flashis the most stable solution. It is 1.11 seconds and is free forever. There are no network problems on domestic platforms.

Cloudflare Workers AIis the biggest surprise this time, with 1.85 seconds of direct connection available. The gpt-oss-120b running on it is the same open source model as Cerebras and Groq. That is to say, you can get the two models that cannot be used due to regional restrictions through Cloudflare.

OpenRouter3.22-second direct connection is available, and it itself is an aggregation gateway, with a key routing to many models.

NVIDIAcan use it but it takes 22 seconds to put offline tasks.

Conversely, if you already have a stable overseas network environment, Cerebras 'speed of 593 milliseconds is the fastest in this batch and is worth keeping.

!

must clarify the boundaries

This test was initiated from the domestic IDC computer room. The network paths and IP segments of computer room outlets and home broadband are completely different, and the operators are also different. The actual results of Jiakuan users may be inconsistent with this, especially those who "can pass by accident", who are likely to perform differently in different network environments.

I can't verify this for you, so I can only suggest you take two minutes to test it yourself. The method is simple and does not require a key:

curl -s -o /dev/null -w "%{http_code}\n" --max-time 10 \
  https://openrouter.ai/api/v1/models

Returning to 200 means that the network is open. Returning to 403 to see if it is error code: 1009. When a timeout is completed, it means that it is completely inaccessible. Change the domain name to the one you want to test.

Let me emphasize the scope again: these 13 entrances are equipped in my own gateway. I chose them because I really use them in production and do not represent all of them on the market. The number of domestic cloud vendors and model platforms far exceeds this number. They do not have accessibility problems to domestic users, so this article does not cover it.

In addition, the registration threshold section is a public compilation of information and is not checked by me one by one. The free quotas and regional policies of each family have been changed quickly. Before officially using them, please follow the instructions on the official website at that time.

The time point is August 18, 2026.