Open LLMs, Real Results: A series on Free-to-Use AI Models

Part 1 - An Introduction to the matter

What will this series be about?

During this series of blog posts, I will do my best to provide my opinion on specific models, determine which use case each one fits, discuss the pros and cons of using that model, and more. I’d like to start with some general notions and first steps, and then deep-dive into different models and particularities of those models. Disclaimer: All opinions on usage and prompts are from personal experience with these models.

What prompted this?

Year 2025 has been a year for new beginnings in my life. Among these are new personal projects. And the frustrations that sometimes come with them. During a rant to a friend about such frustrations (looking at you, configuration issues), he asked me if this is the tipping stone and if I will finally give ChatGPT a chance.

Now, when it comes to technology, especially software, I am one of those people who want to try the new shiny thing as soon as possible and test the limits. However, there is one piece of tech I avoided since it came out: LLMs. I simply did not believe that technology was there. However, I was proven at least partially wrong when I finally tried it for my personal project.

Data privacy and mistakes, or why I avoided LLMs for so long

One of the main issues I had with LLMs is that their data privacy rules were not super clear in the beginning. In an ideal world, such information would be only used to provide the answer and, for longer conversations, to be saved locally on the device and referenced only in the context of that conversation.

However, that has somewhat improved in recent times. For example, OpenAI, Microsoft, and Anthropic do not use the prompts and generated answers for training, while others, like Google, gate this control behind a paywall. Or, in the case of DeepSeek, don’t even mention an opt-out feature.

If we look at how the data is stored, we have the following cases:

• OpenAI - 30 days

• Microsoft - configurable or ZDR (Zero Data Retention)

• Google - 3 years if randomly selected for human review. 3 days with activity tracking off

• Anthropic - unspecified

• DeepSeek - unclear

Then, there is the issue of ownership over the prompts and results. OpenAI, Anthropic, and Microsoft specify that the user has ownership, while Google imposes some restrictions (the data of free-tier users can be accessed by others and is used in training). DeepSeek is again vague, since it uses user data by default for improving the model. There is one thing all the models used as examples above agree on: no copyright guarantee.

The second issue that plagues LLMs is hallucinations and the mistakes resulting from that (how many rocks did I need to eat? One small rock per day?). So, never take what an LLM says for granted.

What is better? Online hosted or self-hosted?

While there are more ways to use an LLM other than Online and Self-hosting, these are probably the most accessible to the average user. So I will focus on these methods at the moment.

I will start with what I would call online hosting. This is the method in which you use the LLM the way it is provided by the creator company, using their apps or websites with an account created and with the settings tuned to your needs.

This is the easiest way, and it is great if you need speed, want to have state-of-the-art outputs, and zero overhead. However, there are drawbacks. You can’t guarantee privacy, you will need permanent network access, and you are always going to worry about your token limit. And sometimes things might just not work, and you will need to wait for some internal service to be back online.

Self-hosting, on the other hand, fixes some of these issues: in terms of privacy and network connection, you are in control. And in some cases, you can even tune the weights. But the machine or machines these run on need to have some competent hardware, the setup time might not be as fast, and it is unlikely that the self-hosted model is state-of-the-art.

So, in conclusion, every person who wants to start using LLMs in their daily life or work will have to decide the best way of accessing them. For me and my project, self-hosting will be the way forward.

What’s next?

Next time, we’ll talk about the first model that I had some more serious contact with: GPT-3.5. I know this is not an open model, but we will use this as a comparison base for the rest. And, additionally, we will discuss the first steps of self-hosting.