Whoever really gives the free quota for domestic large models will find it difficult to get

I flipped through the pricing page of the Intelligent Spectrum Open Platform from beginning to end, and the word "free" appeared 34 times.

Among them, 32 times were in the same column, and the cache storage column said "Free for a limited time".

There are only two models that are free of all four items in the entire row: GLM-4.7-Flash and GLM-4.6V-Flash. Input, output, cache storage, and cache hit are all written "Free" in the four columns. The latter also supports pictures and videos.

I adjusted them four rounds with my own key.

GLM-4.7-Flash: Failed all four times. The first time was a timeout, and the next three times 429 was returned,"The current traffic to this model is too large. Please try again later."

GLM-4.6V-Flash: Success once in four attempts.

For the same key and at the same time, the pricing page did not mark the free GLM-4-Flash. It returned normally all four times, stabilizing at 800 to 960 milliseconds.

Let's be clear, 429 is a current restriction, not permanently unavailable, nor is there any problem with this house. On the contrary, it showsthat the free one is too crowded. But for people who want to prostitute for free, the conclusion is the same: you probably won't use the one marked free.

This incident makes me want to review the "free" offers of several domestic companies. After checking it out, I found that these two words have completely different meanings in each family.

!

Five ## companies, five types of "free"

Intelligence: There are models with free labeling for the entire line, but you can't squeeze in.

These are the two above. The one that can really be used stably is GLM-4-Flash. It is not on the free list, but it is cheap and has never had an accident. In my more than a month of testing, it was the only domestic entrance that had never failed. It accelerated from 2546 milliseconds to 788 milliseconds the day before yesterday.

Tencent Hunyuan: Not free, but the delivery is real.

The official document states that free call quotas will be issued after the service is first opened, and a totalof 1 million tokens will be shared and consumed by the Hunyuan series. The key is the validity period:1 year .

This is the one with the longest validity period I checked. The document also states very bluntly that when the quota is used up, you will have to pay, and there is no permanent free model. Hunyuan-a13b is the cheapest on the list in terms of price, with an input of 0.5 yuan and an output of 2 yuan per million tokens.

Ali Bailian: I gave the most, but only for 90 days.

According to public information, Bailiansends 1 million tokens for each model, and calculates independently for each model. There are many models such as Tongyi, DeepSeek, Kimi, MiniMax, and Smart Spectrum on its platform. If all of them are open, the total amount is quite considerable.

The price is that the validity period is only90 days, and it can only deduct real-time reasoning. Batch calls, context caching, and model tuning cannot be used. I didn't directly confirm this item from the official page. It was compiled according to public information. If I really need to rely on it, please go to the console to check it.

Kimi: Send money, but limit your hand speed.

New users will send a balance of 15 yuan. What really sticks you is not the money, it's the frequency.

I have measured this: 8 requests were sent in a row, andthe fourth request started at 429. The original words returned were request reached organization max RPM: 3. Three times per minute. The description of it in the public information is also "restricted frequency rather than token", which is completely consistent with what I measured.

Therefore, it can be used as a bottom line, and it must fail when used to run batch tasks. There must be more than 20 seconds between requests.

DeepSeek: I don't send it, but I will teach you how to save it.

The official pricing page did not announce the specific amount of the gift quota for new users. But it has one thing that no one else does:half price for off-peak .

The original text says,"The price during idle hours is half of the price during peak hours." Peak hours refer to 9:00 to 12:00 and 14:00 to 18:00 Beijing time from Monday to Friday, and the rest of the time is considered idle.

In other words, if you run the same job at night, the price will be cut in half.

There is also a bigger price difference in the cache: for deepseek-v4-flash input, cache misses are 1.5 yuan per million tokens, and hits are 0.05 yuan. 30 times worse.

!

So the real question is not "Who is free"

Put these five companies together and you will find that "free" is not the same thing at all:

The free version of wisdom score is "bid but can't squeeze in", the free version of mixed essence is "send a year's quantity", the free version of Bailian is "send a lot but expire in 90 days", and Kimi's free version is "have money but only let you adjust it three times per minute." DeepSeek simply refused to give it away and changed it to a discount for off-peak.

The real question is not who is free, but whether you can use the free copy.

This is also the deepest experience I have measured in more than a month. A "who's free" list is useless because there is a gap between the nominal and the actual current limit, a validity period, a frequency upper limit, and whether it can be used for your scene. These things will not be written in the most conspicuous place on the pricing page, so you have to hit them yourself.

!

Based on today's data, I would choose this way

To save trouble and be stable, use GLM-4-Flash. There's no need to grab a free file. It's cheap to begin with, and it's the only domestic portal I've tested that has never been downloaded. Domestic direct connections do not mess with the Internet.

For large quantity and long cycle, choose Tencent Mixed Origin. 1 million tokens plus a one-year validity period is the combination that is most difficult to expire and void among these families.

Focus on running a batch of jobs in the short term and use Ali to refine them. Each model is 1 million yuan. Opening a few more models will cost a lot, but remember that the 90-day clock is ticking, so don't take it and leave it.

Kimi is just the bottom line, not the main force. With the limit of three times per minute, batch tasks will inevitably hit the wall.

If you run a lot without rushing, consider DeepSeek's peak. If you move the task to the evening or weekend, the price will be halved; the price difference for scenarios that can hit the cache will be even greater.

few lines

The 429 of the two models of Zhipu is current limiting. I measured one account, four rounds, and the same time period. It does not mean that they will never be able to adjust, nor does it mean that other people's accounts will encounter the same situation. You might go in at another time.

Tencent Hunyuan and DeepSeek's figures come from their official documents and pricing pages. The part of Ali Bailian is compiled according to public information. I have not confirmed it one by one from the official page. The nature of the source has been marked in the text. There is another saying circulating on the Internet that "Mixed Origin has a completely free model". I couldn't find a corresponding statement in the official documents, so I didn't write it down.

The free policies and current-limit thresholds of each family have been changed quickly. This article is the status of September 3, 2026. Before you really want to place an order, go to the console and check the instructions for the day yourself.