Generative AI Possibilities and Limitations on Personal Offline Usage
Lahti, Matti (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025120934141
https://urn.fi/URN:NBN:fi:amk-2025120934141
Tiivistelmä
The use of generative artificial intelligence has been continuously increasing. With this growth, various concerns have emerged regarding the security and privacy threats associated with generative AI. At the same time, the environmental impact and energy consumption of generative AI have also sparked discussion.
The aim of the study was to assess how well locally run AI solutions on a personal computer compare with the capabilities of cloud services. The research was conducted by executing ten different use case tests both in the cloud and on a computer using various AI models. The resulting findings were then compared against each other.
The study evaluated the quality and usefulness, errorlessness, and scope of the responses generated by different AI models. Additionally, the performance of the AI models was examined. The tests were run on a high‑performance consumer‑grade computer which graphics card was equipped with 96 Gt VRAM for running large AI models.
The results indicated that in text‑based use cases the locally run gpt-oss-120b LLM at a high reasoning level achieved the best overall results. The second best was the GPT‑5 model used by ChatGPT using free plan. Almost alongside the GPT‑5 model, the locally run gpt-oss-120b model at a low reasoning level performed similarly. The gpt‑oss‑20b models fell somewhat behind in the results, especially due to errors that appeared in the translation test.
In addition to high‑quality responses, running AI locally achieved good performance, and the models handling textual data performed at a highly usable level. At a low reasoning level, the execution speed of the local models was comparable to the performance attained by the GPT‑5 model in the ChatGPT service.
For AI‑generated images and videos, cloud services clearly outperformed local solutions. In particular, creating high‑quality videos with a local AI model resulted in poor quality output. In summary, it was concluded that acquiring a dedicated AI computer is sensible if the purchase of a computer is otherwise timely.
The aim of the study was to assess how well locally run AI solutions on a personal computer compare with the capabilities of cloud services. The research was conducted by executing ten different use case tests both in the cloud and on a computer using various AI models. The resulting findings were then compared against each other.
The study evaluated the quality and usefulness, errorlessness, and scope of the responses generated by different AI models. Additionally, the performance of the AI models was examined. The tests were run on a high‑performance consumer‑grade computer which graphics card was equipped with 96 Gt VRAM for running large AI models.
The results indicated that in text‑based use cases the locally run gpt-oss-120b LLM at a high reasoning level achieved the best overall results. The second best was the GPT‑5 model used by ChatGPT using free plan. Almost alongside the GPT‑5 model, the locally run gpt-oss-120b model at a low reasoning level performed similarly. The gpt‑oss‑20b models fell somewhat behind in the results, especially due to errors that appeared in the translation test.
In addition to high‑quality responses, running AI locally achieved good performance, and the models handling textual data performed at a highly usable level. At a low reasoning level, the execution speed of the local models was comparable to the performance attained by the GPT‑5 model in the ChatGPT service.
For AI‑generated images and videos, cloud services clearly outperformed local solutions. In particular, creating high‑quality videos with a local AI model resulted in poor quality output. In summary, it was concluded that acquiring a dedicated AI computer is sensible if the purchase of a computer is otherwise timely.