ChatGPT Replaced: Local AI Put to the Test for 30 Days

Why Local AI is No Longer a Marginal Option

The subscription to ChatGPT Plus, which amounts to $20 per month or $240 per year, has started to weigh on my budget. Using AI daily for writing, programming, and summarizing documents, I was intrigued by rumors that local AI models had reached a performance level sufficient to replace ChatGPT. So, I decided to test this hypothesis.

I installed a local AI on my desktop and my MacBook, using Ollama and Open WebUI for 30 days. I found that these models could effectively handle daily tasks such as writing, summarizing, and brainstorming. In 80% of cases, the results were so close to those of ChatGPT that it was hard to tell the difference.

Models and Performance

Among the models tested, Qwen3 32B proved to be the most effective in terms of quality. For specific tasks, I used models like DeepSeek for complex reasoning and Gemma for quick summaries and Q&A.

The main advantages of using local AI include:

• Privacy, as queries remain on the device
• Cost reduction, especially for batch text processing, where local inference is more economical and faster for repetitive tasks.

Frustrations Encountered

However, the experience was not without frustrations. The main challenges included:

• Failures in multi-step reasoning with long contexts
• Limited or nonexistent understanding of images in most local setups
• Slower response times on CPU for large models
• The time and effort required to properly set up local AI and choose the appropriate models.

Conclusion

In conclusion, while local AI cannot yet completely replace the best cloud models, it is capable of covering the majority of common use cases. A hybrid workflow, combining local AI for most tasks and cloud for the more complex cases, seems to be the most effective solution. Even when returning to cloud usage, the experience has reinforced my privacy concerns, thus changing my approach to AI.