I received 12 free big models and stepped on them for you

I have a principle for connecting AI to the pile of projects at hand: if you can, don't spend money.

It's not cheap. It's just that the number of free models in the past two years has been ridiculously large, and the free quotas of several companies are so large that you can't spend them all on a serious small product. Many people pay for tokens as soon as they come forward, but in fact they haven't done their homework.

In the past period of time, I have selected almost all the large model APIs on the market that can be used for free prostitution, including more than a dozen domestic and foreign companies. Some smell so fragrant after receiving it, and some want to curse after receiving it. Today, I will give this list to you, together with the pits I have stepped on. If you follow suit, you will save me the detours I have taken.

Let's put the conclusion here first. It's enough for those who are anxious to read this picture:

!

Expand and say.

Domestic: These companies that can be directly connected without ladders

** Intelligence GLM-the foundation of my entire chain. **

GLM-4-Flash is the bottom line for all my projects. According to officials, it is permanently free, unlimited in tokens, and only concurrent. Capacity is weak in the free model, but it is better in three points: stability, direct domestic connections, and no sudden charge one day. New user registration also gives a large amount (what I saw was the order of 20 million tokens, subject to the official website), which can be used to unlock a stronger charging model for trial use.

There is one thing worth saying separately: I dare to say that "permanent free will not be capped." It is said that in China, there are only Smart Music and Tencent Cloud, and the free services of other companies are basically limited time activities or gifts from newcomers. So even if you mainly use something else, it is recommended to put the smart spectrum at the bottom as a fuse.

** Volcano Engine-a lot of money, but it was also the one I suffered the worst. **

It has a free quota of millions of tokens every day (I use the 2 million per day limit), which is large enough to ensure your fill. But this house is free of charge. I will leave a separate section after the setback I made on it. Please remember one sentence first: Don't make random adjustments, you will deduct money if you adjust the wrong access method.

** Ali Bailian/Tongyi Thousand Questions-You can type in both Chinese and code. **

The Qwen series is one of the most capable open source models in China. It is good in Chinese and coding, and individual developers have long-term free quotas. One detail should be noted: some of its free quotas have a validity period (what I see is a three-month period), and will expire after expiration, so don't hoard them and not use them.

** Silicon-based flow-an "open source model aggregation supermarket". **

It helps you unify a bunch of open source models such as DeepSeek and Qwen into one interface, and new users will send quotas, saving yourself the trouble of deployment. Suitable for you to compare back and forth between several open source models.

** Magic ModelScope-Multimodal is more suitable for playing in the domestic free game. **

Ali's model community has free reasoning quotas every day, and multimodal such as graphics and text are given generously.

**Kimi--The strategy is quite special, regardless of the amount or frequency. **

It means that you can use it to slowly gnaw through ultra-long documents. As long as you are not in a hurry, the amount of money you can pay for is actually quite considerable. It is suitable for doing long-term summaries that do not pursue real-time work.

Overseas: Strong and fast, but both must have ladders

These domestic companies cannot be directly connected, so they have to solve the network themselves. But they have things that they can't give domestically for the time being-either ridiculously large amounts or ridiculously fast.

**Google Gemini-the one with the highest free limit. **

According to officials, the Gemini Flash file is about thousands of requests a day and a million tokens per minute. It does not need to bind a credit card or expire. Among the ones I took, it ranked first in terms of the amount of white prostitutes. The only thing you have to accept is that free files may be used to train models, and don't put sensitive data in it.

**Cerebras-Unreasonably fast, but free catalogs float. **

It uses wafer-level chips specifically for reasoning. Short requests often produce results within one or two seconds. There are still 1 million tokens a day, and no credit card is needed. But there is one problem with it: the models that can be used for free change just like they want-I've seen its free catalog cut from a dozen to two overnight. So never write the name of a specific model into code.

**Groq-the king of speed, but the token is stingy. **

It's a little nice to have a short request coming quickly. The price is that the tokens are tight. The flagship 70B model costs about 100,000 tokens a day, and it will be exhausted after more than a hundred calls a day. Suitable for high-frequency short requests, don't feed it to long-term texts.

**GitHub Models-one of the few places where you can encounter GPT-4.1 for free. **

You can use it with a GitHub account, and you can tune in to first-line closed-source models such as GPT-4.1 and GPT-4o-mini for free. The amount is not large (dozens of times a day), but the advantage is that the goods are good. I use it to specialize in low-frequency tasks with high quality requirements.

**Mistral-In this European house, free files also give a large amount of money every month, which is suitable for use as a backup. **

!

5 pits I stepped on for you

Anyone can make a list, and pits are really valuable. I spent real money and even locked my account to remember the following few.

!

** Keng One: "Automatic transfer to payment when the free quota is used up"-this is really a way to deduct money. **

The most ruthless one. On some platforms, you nominally use free credit, but the access method you call is "use it for free first, and automatically charge it by token when you use it." As soon as the limit is crossed, the money will start to be deducted, and you may not find it immediately. Because I adjusted the wrong access point to the volcano, my account owed a fee-what's more, as soon as the account owed, it locked all the models in my name, including those that were originally free. Before opening it, be sure to make sure that you are in the "purely free, exhausted and rejected" position, not the "free transfer to later payment" position. I also received a text message from Baidu saying "Post-payment has stopped". The logic is similar, it's all about this trap.

** Pit 2: Most "free" does not mean "permanent". **

As I said before, there are only two companies in China that dare to say that there is no cap for permanent free. The vast majority of the rest are "gifts from new users" or "time-limited activities", which are used up or are invalid after expiration (some files in Alibaba expire in 3 months). Don't make plans with time limits as permanent.

** Pit 3: Current restriction blocks a hard wall, and you will hit 429. **

Almost all free credits have per-minute and daily calls and token limits. If you hit any one, the request will be directly rejected (429 error). In front of the Groq 70B, it sounds like a lot, but it's gone after running hundreds of times. To make a serious application, retry and downgrade must be written in advance.

** Pit 4: Overseas companies need ladders, and current limits are based on accounts not key. **

Groq, Gemini, Cerebras, GitHub, and Mistral cannot be directly connected in China. And don't bother with the idea of "registering more keys to bypass current flow"-these companies basically calculate the limit based on accounts or projects. Opening more keys is useless, and opening more numbers may also be blocked in violation of the terms.

** Pit 5: Free model, remove it from the shelves as you say. **

Cerebras 'free catalog is cut from a dozen to two overnight. This is the norm. For free things, the platform has no obligation to give you stability. So if you rely on it, be sure to do fallback: the main model hangs and automatically cut the next one, don't write a single point dead.

, so don't choose one, build a chain

After taking a round, my biggest experience is: In terms of free large models, the correct posture is not "picking the best one", but "building a chain that will relay automatically."

What I have now is probably this logic:

Put a permanent free fuse (Smart Spectrum GLM-4-Flash) at the bottom of the - -the front is all hung, at least it is still there, and the service will not be interrupted; The main force of - intermediate amplification quota (Volcano, Gemini, etc.) carries daily life;

• requires fast high-frequency activities and hand them over to Cerebras and Groq;
• low-frequency tasks with high quality requirements, use GPT-4.1 on GitHub; Each layer of the - is set to "automatically cut the next one if this hangs".

With this set, the model cost of a small and medium-sized project can be close to zero. And it is more risk-resistant than hanging on a paid API-after all, free things change as they say, and you have to let the system know how to do it.

Last word of truth: the free policy is changing very quickly, and every number in this article may be inaccurate in two months. Before picking up, take two minutes to go to the corresponding official website to confirm the current limit-this is the most cost-effective two minutes.