I started Thoth with the free models from ArliAI, but somewhere along the way I learned that Gemini has a free tier, too. I never got around to updating the variable naming, but right now, Thoth works with both sets of models. Eventually, I might add Hugging Face API, too, if I can find time and figure out how. I've seen claims that they have a free tier, but I haven't been able to confirm that so far.

The free tier from ArliAI currently offers the Qwen3-14B family, but those results are currently unimpressive. The gemini family which is available for free from Google produces far superior results right now (You can see those results at the @thoth.test account, and you can see a Qwen3-14B run here.).

The main point of Thoth isn't the AI analysis - that's just an enabler. The main point is to provide an alternative to the spammy delegation bots that are currently flooding the social ecosystem here with pollution.

From the original post, I agree with @moecki that there's probably not a free way to do the AI detection at the scale needed to evaluate every post. Eventually I hope to plug both plagiarism detection and AI detection into Thoth, but I'm thinking that both will probably require paid services. The only reason Thoth operates for free is that I do a lot of screening using traditional filtering techniques before the LLM ever sees anything.

Thoth's prompt does direct it to filter out AI and plagiarized content, but I doubt very much if the LLM really has the capability to do that.

As you note, we could roll our own AI detection, but I think that would require more compute & storage resources than I have available to me, and I think it would probably just be cheaper to get a paid subscription to an API from a service that already does it.