13 free big models, 4 of which were hung up in a month

There is a list of free large models hanging on my server, with 13 entrances, which were calibrated one by one on July 12-which quota, how many times per minute can be sent, and how slow it is, were all tested one by one at that time. It was not copied.

I ran this list over again yesterday.

Of the 13, 4 were killed.

Less than a month.

!

first mentioned the most unexpected one: GitHub's free model, it's gone

There are two entries in the list that go to GitHub Models-GPT-4.1 and GPT-4o-mini. When tested on July 12, GPT-4.1 was about 50 times per day, and 4o-mini was about 150 times per day. It was a rare and serious channel for free sex with the OpenAI model.

I adjusted it again yesterday and returned this:

github_models_retirement_brownout:
GitHub Models is temporarily unavailable
as part of a scheduled retirement

After checking the official announcement, GitHub Models will be completely retired on July 30, 2026 . Playground, model catalog, inference API, and BYOK are all offline. Retirement is divided into three steps: on June 16, new customers were stopped, on July 1, it was announced that the total shutdown would be completed on July 30, and a brownout was conducted on July 16 and July 23 (short-term break exercise to remind you to move quickly).

The key is that this last strike does not exempt old users. Even if you have been using it and the amount is normal, it will still stop on July 30.

My list was calibrated on July 12. Eighteen days later it died.

The other three died differently

ModelScope's DeepSeek-V4-Flash, a month ago, it was 2000 times per day, but now it reports:

400 Model id : deepseek-ai/DeepSeek-V4-Flash,
has no provider supported

It's not that the current limit is current, but that this model ID no longer has a supplier on the platform. The model itself was removed and the interface was still there.

OpenRouter's Llama-3.3-70B free, reported as:

404 This model is unavailable for free.
The paid version is available now

The free version will be removed and the paid version will continue to be sold. This is the most typical way to die-not shutting down, but withdrawing free files.

SambaNova reported PAYMENT_METHOD_REQUIRED, requiring that the payment method be tied first. However, it died in July, and the list was already marked "dead" at that time. This time, it was just confirmed that it was still dead.

So strictly calculated, There were four new entrances posted this month: two on GitHub, one on ModelScope, and one on OpenRouter.

nine alive, 30 times faster

After saying the dead, talk about what can still be used. Adjust the same question and the same parameters one by one:

!

model	Single time consuming	Concurrent 5
Cerebras oss-120b	576ms	5/5
Groq oss-120b	671ms	5/5
Gemini Flash-Lite	1267ms	5/5
Smart Spectrum GLM-4-Flash	1543ms	5/5
Mistral Large	2668ms	4/5
NVIDIA Nemotron-49B	4842ms	5/5
Smart Spectrum GLM-4.7-Flash	6290ms	returns an empty
Kimi k2.6	12284ms	5/5
Volcano GLM-5.2	17574ms	5/5

The fastest Cerebras is 576 milliseconds, and the slowest volcano, GLM-5.2, is 17.5 seconds. 30 times worse.

This gap is fatal because they are arranged side by side in the list and all seem to be called "free." But for the 576-millisecond one, the user waits on the page and comes out; for the 17.5-second one, any scene where someone is waiting in front of it cannot be used-the user has long thought it was stuck and left. It is only suitable for batch processing at night.

In the concurrent test, only Mistral hung one , and one of the five concurrent requests received 429. This is in line with its "2 times per minute" limit: it gives a generous daily quota (1 billion tokens per month), but the tap is turned on extremely small. Large quotas and being able to collect money are two completely different things . This is one of the misunderstandings I have seen most people fall into.

There is another trap to mention separately: Intelligence GLM-4.7-Flash has been switched on, but the content is empty. It's not that the model is broken, it's that it belongs to the thinking model. max_tokens is the total budget of "thinking + answering", and thinking itself consumes more than a thousand tokens. If you give 500, it's all used up by thinking, and the answer will be written without budget. Return an empty string, finish_reason is length. Only when you give more than 1500 can you have stable content.

is a trap for not reporting errors: For the same model, if the call method is wrong, it changes from free to charging

Volcanoes GLM-5.2 will be discussed separately because it is the most prone place to lose money on this list.

There is an iron rule in the list: You must use the EP ID of the "Authorized Access Point"(like ep-m-2026...) to adjust it. Only by walking through this channel can you count as a collaboration reward and be free. If you want to make it easier and use the Model ID glm-5-2-260617 directly, the backend clearly states that you will pay by using Token.

The same model, the same key, and the same interface, but the model field in the request body is different. One does not cost money, and the other starts charging.

I retested the EP ID this time, so the above 17.5 seconds is the result of the free channel.

The most sinister thing about this type of pit is that it does not report errors. You can still adjust and get the results. There is no difference in experience, but the bill is waiting for you. By the time they were discovered, they had usually run tens of thousands of times.

By the way, the bean buns, DeepSeek, and image models from the same house are all paid by Token, so don't add them to the free pool-and once the account is in arrears, will also lock the free one .

In the ## limit column, I did not retest this time

The boundary that needs to be clarified: This time I re-ran the survival, delay, and concurrent , and there is no retest quota-it takes a lot of calls to truly run through the daily quota of each household, and the cost is too high.

The following amount was recorded on July 12 and can only be used as a reference now:

model	The free amount recorded at the time
Volcano GLM-5.2	Daily limit of one million, exhausted refuse, no charge
Mistral Large	1 billion tokens/month, but 2 times/minute
Cerebras oss-120b	1 million tokens/day, about 30 times/minute
Gemini Flash-Lite	About 1500 times/day
NVIDIA Nemotron	40 times/minute, no daily limit
Groq oss-120b	1000 times/day, but the maximum is 100,000 tokens/day
Smart Spectrum GLM-4-Flash	Permanent free without cap
Kimi k2.6	Taking the balance, there were less than ten yuan left at that time

Considering that four entrances were hung in a month, how many of these numbers are still accurate now? I can't guarantee. Use it as a reference, not as a basis.

How can I arrange ## myself?

To be fast and carry concurrent , use Cerebras or Groq. Half a second to seven hundred milliseconds, all five passes are passed simultaneously. These two are the only ones that can be put in front of the user.

Domestic direct connections do not want to use agents , use Smart Spectrum GLM-4-Flash. 1.5 Seconds, all concurrent passes, permanent free without cap. It has average capabilities, but it never loses its chain. It is the last fuse on all my links.

It takes a long context, energy, etc. before it turns to volcano GLM-5.2. A 17-second response can only accept offline batch processing.

Use Mistral alone to control the frequency , don't issue it concurrently, just queue up honestly.

Remember to give max_tokens when thinking about the model, otherwise it will be white.

!

The shelf life of ## list is about a few weeks

When I wrote this, I was actually a little hesitant whether to post it.

Because at this decay rate-four per month-the table in this article will probably have a few lines wrong after a month. I give you a list, and it starts to rot when you turn around, which is a bit absurd in itself.

But on the other hand, this is exactly the one thing that should be said: Most of the "2026 Free AI Model Complete Book" on the Internet were copied a few months ago, but they were not tested when they were copied, and no one went back and retested. The one you collect seems to have 20 to 30 brands, but in fact, half of them may be dead. The retirement of GitHub Models is still recommended in countless "Free Free Whoring Guides".

So what is more useful than this list is Don't believe any undated list , including this one. Its date is August 10, 2026.

Before you really want to use it, it's safest to spend two minutes to run through it yourself-send the simplest requests one by one to see who still responds. The entire retest took only ten minutes.