A Brief Discussion - Testing Mainstream Open-Source Large Models on Laptops

🌐 中文

English

This test primarily involves common local deployment model testing on laptops.

Introduction

I saw this message in my social media feed: This sparked an idea: I wanted to test the capabilities of current mainstream open-source LLMs under resource-constrained conditions on a laptop, using the question from the social media post: How many numbers between 0 and 1000 contain exactly one digit ‘5’?

Test Results

1. ollama3.2:3b
2. qwen2.5:7b
3. ollama3.1:8b
4. gemma2:9b
5. phi3.5:8b This model is hard to describe; it doesn’t give the answer directly. Finally, I found this passage:
   It seems to give the conclusion of ‘only one’, which is also a kind of ‘only one’.
6. mistral:7b

It can be seen that none of these local models answered correctly.

Summary

Currently, deploying mainstream open-source large models with many parameters locally requires significant resources. Therefore, under resource-constrained conditions, the mathematical capabilities of these models are still relatively weak. Thus, AI open-source model practitioners need to explore more in areas including architecture, parameters, and algorithms to develop useful open-source large models that can be deployed on laptops, allowing AI to benefit all of humanity.