Of the 21 free large model entrances, only 9 are actually available
The article six days ago said that 4 out of 13 entrances were hung, and a few more have been added in the past few days. I have adjusted all 21 entrances one by one: 9 are usable, 5 are current-limited, and 7 are dead. But what is more useful is not these numbers, but that the same model is now often hung on several platforms, with completely different usability. The three entrances of gpt-oss-120b have two connections and one limit. The most counter-intuitive one is GLM-4.7. Smart Spectrum has its own current limit, but the one managed by Cerebras can be used instead.
Six days ago, I wrote an article saying that 4 of the 13 free large model entrances were hung up in one month. Several more entrances have been added in the past few days, and I have adjusted all 21 entrances one by one.
The result is: 9 are available, 5 are currently limiting, and 7 are dead.
But it's not these numbers that are really useful this time. It was when I was checking the data that I realized that I almost sentenced a person to death.
First look at the list
The same question and the same parameters, measured on August 17, 2026:
Available (9)
| inlet | time-consuming |
|---|---|
| Cerebras oss-120b | 593ms |
| NVIDIA Ultra-550B | 1343ms |
| Smart Spectrum GLM-4-Flash | 1590ms |
| Mistral Large | 1687ms |
| Cloudflare oss-120b | 1764ms |
| GLM-4.7 on Cerebras | 1938ms |
| OpenRouter Nemotron | 2358ms |
| NVIDIA Nemotron-49B | 4058ms |
| Kimi k2.6 | 10311ms |
Current limiting (5): GLM-5.2 on Mistral, Gemini Flash-Lite, NVIDIA MiniMax-M3, Groq oss-120b, Smart Spectrum GLM-4.7-Flash. I tried each of these five three times every 12 seconds, but they were still restricted.
Already dead (7): Volcano GLM-5.2 (account arrears), GitHub GPT-4.1 and GPT-4o-mini (service retired), Groq's Llama-3.3- 70B (model ID no longer exists), OpenRouter's Llama-3.3 (free file withdrawn), ModelScope's DeepSeek-V4 (model removed), SambaNova (required to bind payment method).
Cloudflare is a new face in this batch, gpt-oss-120b on Workers AI, 1.7 seconds, and works.
The one I almost wrote wrong
After the first round of scanning, GLM-4.7 on Cerebras was ruled a failure by my script. When I was about to write it in the "hung up" column, I took a look at the original return and found that it actually succeeded.
The problem lies in my decision logic: I look at whether there is content in the content field. This model is an inference model. It puts the 1358-character thinking process in the reasoning field. There are only 40 words in content. In addition, the max_tokens I gave in the first round was too small, and the text was squeezed out, so content was empty.
I sentenced a living model to death because of its different return structure.
There are two useful conclusions to this matter.
For people writing code: To judge whether a call is successful or not, it is not enough to just look at content. The inference model may put the content in reasoning or reasoning_content. The thinking model I wrote on August 10 will eat up the token budget. Now I'll add one more: it may also put content in fields that you haven't read.
For those who watch the assessment: Any list of "what can and cannot be used", including mine, depends on how the person taking the assessment determines that "it can be used." If the judgment logic is weak, the conclusion is short of a model.
QKPFX2 One thing QK is more useful than the list
Rearrange these 21 entrances according to "Which model is actually running" and you will see a structure that you didn't notice before.
The same model is now often hung on several platforms, and the usability of each platform is completely different:
Gpt-oss-120bhas three entrances. Cerebras is available for 593 milliseconds, Cloudflare is available for 1764 milliseconds, and Groq is current limited.
The Nemotron-3-Ultra-550Bhas two entrances, both NVIDIA and OpenRouter.
GLM-4.7has two entrances. What is available on Cerebras is available,but its own is limiting current.
GLM-5.2has two entrances. The volcano owes money and stops serving, and Mistral has current restrictions, so neither of them can be used.
Llama-3.3- 70Bhas three entrances. The Groq model ID is invalid, the OpenRouter free file is withdrawn, and SambaNova wants to bind the card, so everyone is annihilated.
The most counter-intuitive thing in this set of comparisons is GLM-4.7: Wisdom is the owner of this model, but today its own free entrance is restricted, while the one hosted by Cerebras is available.
Therefore, the question "Can a certain model be paid today?" is a wrong question. The question should be,"Where can this model be used to prostitute for free?"
This means how should you match it?
If you hang free models in your project, the most practical suggestion given by this data is: the same model should be equipped with multiple platform entrances.
The reason is straightforward. Among the above five groups, if a model with only one entry is hung, you will have no backup; while gpt-oss-120b has three entrances. Today, Groq is restricted, Cerebras and Cloudflare can still top it, and your program won't even notice any abnormalities.
This is more effective than "matching several different models." Because different models mean different output styles, different parameters, and different return structures, it is very likely that the code will have to be changed when switching in the past; while the same model changes platforms, the request body will basically not be moved.
Specific to the data this time, I will arrange it as follows:
gpt-oss-120b takes Cerebras and prepares Cloudflare. Cerebras 593 milliseconds is the fastest in the game, and Cloudflare is currently stable as a new entrance. The entrance to Groq remains. It is normal for it to restrict current but it recovers quickly.
If you want to connect directly in China without using agents, Smart Spectrum GLM-4-Flash is still the most reliable way to provide a solution. 1590 milliseconds, permanently free. I calculated in the previous article that its success rate among 25,000 real calls was 98.2%.
Kimi recovered. In my production log in early August, its success rate was only 0.3%, and it was almost completely useless. This time, the energy test returned normally. Although it was slow in 10 seconds, it was alive.
Stop spending time on Llama-3.3- 70B. All three entrances are gone.
few sentences boundary
This is a snapshot of an account and a point in time, August 17, 2026. The five current-limiting items may be ready tomorrow, and the nine that can be used may be limited tomorrow. The shelf life of such lists is measured in days.
The distinction between current limit and dead is based on the error content: those reporting insufficient balance, service retirement, model does not exist, and demand payment are classified as dead; those reporting 429, rate limit, and excessive traffic are classified as traffic limit, and each was retried three times. This division has subjective elements. For example, the volcano is in arrears and can be resurrected after recharging. Strictly speaking, it is not "dead" but "my account is in arrears."
In addition, I only tested whether I can return the result, not the quality, nor did I run through the quota to see the upper limit. Being able to be adjusted and easy to use are two things.
If you want to check it yourself, the judgment logic should remember to read content and reasoning at the same time. max_tokens should be given above 1200. Don't judge the living ones as dead ones like I did in the first round.