Free big model, which is the smartest? I ranked the problem
Since they are all free, of course we collect the smartest ones. I took three reasoning questions with the only correct solution and accidentally fell into the pit and adjusted the seating order of each free large model: Kimi K3, Cerebras, and Groq with full scores in all pairs;Gemini machine questions; the most powerful GLM-4-Flash has three errors and two errors at the bottom. Core: Speed and cleverness cannot be achieved at the same time.
Horizontal evaluation of which one can use and how much wool one can collect. Today, the most practical one is collected: Since they are all free, which one is the smartest?
You have to know who to pick up before picking up. I took a few difficult questions that could widen the gap, adjusted each free model, and arranged a seating order. They are all inferences with only correct solutions, but they fall into the trap if you are accidentally careful.
Let's see the results first.
There are three who get full marks: Kimi K3, Cerebras 'oss-120b, and Groq's oss-120b. All three deduction questions were correct, but none of them failed. To do work that requires brain use, it is reliable to collect one of these three.
What's interesting is that the one at the bottom is everyone's favorite GLM-4-Flash. It is the king of wool harvesting. It is unlimited and free forever. I recommended it as the main force in the previous few articles. But when it comes to an inference question, it makes three mistakes and two mistakes.
First question: Sally has 3 brothers, and each brother has 2 sisters. How many sisters does Sally have? The correct answer is 1. There are only two girls in the family, Sally and her sister. GLM-4-Flash circles around and creates a 4. Kimi K3, Cerebras, Groq, and Gemini all answered 1 steadily.
Third question: 5 machines make 5 parts in 5 minutes. How long does it take for 100 machines to make 100 parts? It takes 5 minutes to correct the problem, one for each machine for 5 minutes, and doing it at the same time for 100 machines, which still takes 5 minutes. This question was an intuitive trap. GLM-4-Flash answered it for 1 minute, Gemini answered it for 100 minutes, and both failed. Kimi K3, who can reason, as well as Cerebras and Groq, all got it right in 5 minutes.
String together the actual measurements of these articles, and a complete portrait of GLM-4-Flash, the "king of wool harvesting", comes out: it is the most capable of harvesting (no current restriction, permanent free), but it is also the least trustworthy (loves to make up nonsense, the weakest reasoning). There is a reason why it is cheap.
On the other hand, Kimi K3 ranks first in ability. It was the only one that didn't make up the previous illusion test, but it was slow and there was a risk that the free window would close at any time.
So if you collect free big models, don't just get them all into black, it depends on the live points:
Brainless, high-frequency work, translate a short sentence, change a format, run a batch, and use GLM-4-Flash to open up the collection. Anyway, it is simple, it does not limit current, so you can collect as much as you can.
You need to use your brain, reason, and ask it to be trustworthy. Don't use those fast and cheap ones. Lightweight versions such as GLM-4-Flash and Gemini Flash-Lite are the most likely to overturn; those who are willing to calculate power such as Kimi K3 or Cerebras and Groq are smart and not fabricated, at the expense of being slower or limited.
At the free level,"fast" and "smart" often cannot be achieved at the same time. Find out which one is smart, which one is fast, and which one can collect freely, and then you can collect more and more accurately by pressing the lively party.