Problem No. 1: Small llms stupid
The latest LLMS Open often boasts of major standard improvements, and this was definitely the case with Deepseek-R1, which approached Openai’s O1 in some criteria.
But the model you run on the Windows laptop is not the same as it records high marks. It is a smaller and more condensed model – the smaller versions of large language models are not very smart.
Look just to what happened when I asked Deepseek-R1-Lama-8B how the chicken is crossed:
Matt Smith / Mabek
This simple question – and I answer the street llm – passes how smaller models can easily start from the bars. They often fail to notice the context or pick up the nuances that should look clear.
In fact, recent research indicates that large language models are less intelligent with thinking capabilities are vulnerable to such breakdowns. I recently wrote about the issue of thinking about the models of thinking about artificial intelligence and how to increase mathematical costs.
I will admit that the example of chicken is ridiculous. What about we tried a more practical task? Like a simple web coding in HTML. I have created a fictional CV using Sonnet Claude 3.7, then I asked QWEN2.5-7B-Instruction Create a HTML website based on the CV.
The results were far from the wonderful:

Matt Smith / Mabek
In order to be fair, it is better than I can create if I sat on a computer without an internet connection and asked me to encode a similar web. However, I don’t think most people want to use this CV to represent themselves online.
A larger and smarter model can be born, such as Claude 3.7 Sonnet of the human, high -quality web site. I can still criticize them, but my problems will be more accurate and less relationship with blatant defects. Unlike the QWEN product, I expect that many people will be happy to use the website that has been created to represent themselves online.
For me, this is not speculation. This is actually what happened. Several months ago, I abandoned WordPress and moved to a simple HTML website coded by Claude 3.5 Sonnet.
adxpro.online