Free large models were tested continuously for 22 days, and only 3 were on full attendance
From August 17 to September 8, in 22 days, I tested the same batch of free large model entrances five times, with the same script and the same question. Only 3 of the 21 entrances were available all 5 times. If you stack up the five records, you can see what you can't see in a single test: the three common points that survived were "resident free files" rather than "free quotas"; among the ones that died, one type would slow down first before death, and the Nemotron-49B officially stopped serving the next day after it was 5 times slower; and the error reporting of the current restriction basically came back to life on its own. Groq jumped 4 times in 22 days, and won the first place directly in both resurrections.
From August 17 to September 8, in 22 days, I completely tested the entrance of the same batch of free large models five times. The same script, the same question, the same parameters.
Of the 21 entrances, only 3 were available 5 times.
This article doesn't talk about which one can use today. I've written that kind of list several times, and the shelf life is only one week. This time, we talk about what we can see after stacking the five records together: what kind of free entrances live long, what kind of free entrances will suddenly disappear, and whether there are signs before death.
full attendance 3
Smart Spectrum GLM-4-Flash, Cloudflare's oss-120b, Kimi k2.6.
| inlet | 8-17 | 8-25 | 9-02 | 9-04 | 9-08 |
|---|---|---|---|---|---|
| Smart Spectrum GLM-4-Flash | 3827ms | 2546ms | 788ms | 1634ms | 1040ms |
| Cloudflare oss-120b | 2109ms | 4205ms | 2108ms | 2512ms | 1620ms |
| Kimi k2.6 | 10541ms | 11716ms | 8970ms | 10877ms | 3910ms |
The speed fluctuates, and the intelligence spectrum fluctuates between 788 and 3827. Kimi has always been slow but this time it is much faster. But they never dropped the line once in five times.
I thought it was just luck at first. When I pulled out the dead batch and compared it, I found that these three had one thing in common: they all gave "resident free files", not "free quotas."
This difference is fatal.
Resident free files are a location that has been open on the platform for a long time to attract people in.It doesn't promise how much you can use, but this gear is always there.
The free quota is to give you a sum until you use it up.
The dead batch were basically the latter:
Cerebras gave a free credit. On August 17, it was the fastest 593 milliseconds in the game. I clearly recommended it as the first choice at that time. Test again 8 days later, Payment required to access this resource. The limit is up.
Mistral follows the subscription level, This model is not available in your subscription tier.
SambaNova is more direct and requires that the payment method be tied before the adjustment is made.
So when choosing a free entrance, first look at which type it is. The quota type is a consumable item that you just haven't used up yet; the resident file is the platform's guest seat, and it has the motivation to keep it open. This judgment is much more useful than seeing who is faster.
will slow down before it dies
This is the one I found out from this review that I didn't expect myself.
NVIDIA's Nemotron-49B is the cleanest sample:
August 17, 2534 milliseconds, normal. August 25, 15574 milliseconds, which was five times slower, but it still worked. August 26, official EOL, return with a clear date. Then test again, 410 Gone.
There was only one day between slowing down and stopping service.
I saw it at 15574 milliseconds and thought it was a bad network that day and didn't think much about it. Looking back now, this is what it looked like before it died.
What's even worse is that there are now two entrances walking on the same curve:
| inlet | 8-17 | 8-25 | 9-02 | 9-04 | 9-08 |
|---|---|---|---|---|---|
| NVIDIA Ultra-550B | 2863ms | 1377ms | 7007ms | 8257ms | 35328ms |
| OpenRouter Nemotron | 2318ms | 1437ms | 2629ms | 9873ms | 37407ms |
It slowed down for two consecutive periods, and the last period rushed directly to more than 35 seconds.
These two are two different entrances to the same model that slow down together. It shows that the problem lies with the model itself, not the routing or current restriction of a certain home.
They haven't reported a single mistake so far.You adjust it, it comes back to you, but it just takes half a minute. The interface level is completely healthy, monitoring will not sound, and the business has long been abandoned.
It has slowed down for two consecutive periods, so it is time to prepare for the back path. This is the only way to die with signs, but it won't give you any wrong results and can only be discovered by regular measurements.
three ways to die, whether you can come back to life is completely different
Don't rush to cross it out when you see an error reported, classify it first.
| way to die | What does a wrong report look like? | Has he been resurrected in 22 days? |
|---|---|---|
| current-limiting | Rate limit reached、429 | have lived for |
| Limit/Pay Wall | Payment required, insufficient subscription level | None came back |
| retired type | end of life、410 Gone | would not come back |
Groq's journey is most telling:
The current restriction was imposed on August 17, and I ruled it suspended. Available on August 25, 545 milliseconds, the fastest game. On September 2, the current restriction was imposed again, and I ruled it dead again. Available on September 4th. Available on September 8, 579 milliseconds, returning to first place.
In 22 days, he jumped sideways 4 times and finally survived. Moreover, he won first place directly twice.
Smart Vision's GLM-4.7-Flash is more extreme. The current limit was imposed on August 17, the current limit was imposed on August 25, and the 120-second timeout attempt on September 2 were all defeated. I put it on the death list for three consecutive periods. It suddenly opened on September 4th, taking 75 seconds. September 8th, 19 seconds.
After three consecutive periods, he is back to life and is getting faster.
Conversely, Cerebras did not come back for 22 days after hitting the pay wall, and Mistral did not come back after falling into the subscription level. These two categories will not be good on their own.
So: Just wait for current limit, change pay wall homes, and delete EOL directly.Treating current restriction as death will cost you throwing away the fastest one in vain. I have already thrown it away twice.
The first place in ## has changed four times, and the first three have died
| date | first place | time-consuming | How did the previous term go down? |
|---|---|---|---|
| 8-17 | Cerebras | 593ms | first measurement benchmark |
| 8-25 | Groq | 545ms | Cerebras hits paywall |
| 9-02 | Smart Spectrum GLM-4-Flash | 788ms | Groq current limiting |
| 9-04 | Groq | 598ms | Intelligent spectrum is slow to 1634 ms |
| 9-08 | Groq | 579ms | defending |
It was changed 4 times in 22 days, with an average of 5 and a half days.
This rhythm means that any "who's the fastest" conclusion has a shelf life of a few days, including those I wrote myself.If you really want to use it, don't believe in rankings, believe in a list that can be automatically switched.
four items can be used directly
Relying on resident free files and not on free quotas. The quota is a consumable item, you just haven't used it up yet; the resident file is the platform's guest seat, and it has the motivation to keep it open. All three people on full attendance are the latter.
The same model is equipped with multiple platform entrances. gpt-oss-120b is available in Cerebras, Groq, and Cloudflare. At least one of them has been alive in the past 22 days, and never all of them have been hung up at the same time. There is basically no need to change the request body for changing the platform, and the code for changing the model is required, and the cost is far from that.
Prepare for the back road when it slows down for two consecutive periods. Performance degradation is the only way to die with signs, and the price is that it does not report errors. Add a timeout to the call. If it exceeds, cut the next one without waiting for it to return you an error.
Current limiting is not death, retreat and retry. Play again every few dozen seconds for a few more rounds. Groq won first place immediately after being resurrected twice, and I almost missed it twice.
few sentences boundary
This is the record of an account at 5 time points. If the amount of the free layer is calculated based on the account, your result will be different from mine, and the current restriction is especially obvious.
There are two categories mixed in "unavailable". The service itself is gone, such as the retirement of GitHub Models, and the clear EOL date of Nemotron-49B, which is certain; there are also cases where I can't use this account, such as arrears, subscription levels, and expiration of quotas. Changing the account may be good. I tried to write it separately in the main text.
Since September 8, I have set the available threshold at 35 seconds. There was no such threshold in previous issues. Therefore, the Ultra-550B and OpenRouter Nemotron were ruled unusable in the final issue. Strictly speaking, they were "too slow to be used" and not "unable to adjust."
I only tested whether it could return normally, did not measure the output quality, and did not run through the quota to see how much money I could pay for.
These 21 entrances are equipped in my own gateway, not a full-market census. There must be something on the market that I haven't detected.