I will be explaining Discourse AI integration to the Team next week. We already have some api keys, BUT, they are all for PAID accounts.
I am sure I will be asked; Are there any Features that can be integrated that can be used for Free ? This is for a school and we are sure students may want to use everything, but the costs will be out of this world.
Thanks
No. It’s not free.
Actually, llama3 is free, but best I can tell, it takes a $300,000 computer to run it.
If you have some budget, you could set it such that you pay that much monthly to whatever service you want to pay for and when it’s used up for the month, it’s used up for the month. You’d try to set up limits so that it wouldn’t happen the first week (or day). It would be complicated to set it up such that teachers who wanted to use it as part of class could count on it.
As a school maybe you can signup for free educational credits on services such as AWS, Azure or GCP and use that to power the Discourse AI features.
Thanks that’s what I planned to tell them. It’s not Free. The plan for now is to continue to use local instances of Gpt4All and LmStudio
For what it’s worth, you can run the 70b version of Llama3 within 48GB VRAM which you can source relatively easily from a pair of used Nvidia RTX 3090’s from eBay for about ~$750 each. Building out the rest of a system to support that would land in the ballpark of ~$3000-ish I expect. 
That’s a factor of 100 improvement! I can imagine a school having that much money, but I have a pretty good imagination.
Thanks very much for that. I mostly don’t understand the hardware requirements.
You can look at this in a different way.
Why not pilot the use of AI and watch the cost dashboards for a month? Then decide if you want to continue.
Just in the last 12 months the cost of capable models has come down A LOT.
OpenAI’s gpt4o-mini is very capable and very cheap now (but it’s all relative.)
Well, mini has failed me everytime. But that is worth of another topic somewhere else.
Yes, it may depend on the language too.
Gemini Flash is very cheap and DeepSeek API is so cheap it is almost free.
If you need to run locally, you can build a machine for <$1000 that can run Llama models. I built a 4xP100 machine for $1000 that has 64GB VRAM. But a 2xP40 machine with 48GB VRAM would be enough to run 70B Llama and can be built for $600 if you buy 2nd hand parts.
For the ultimate in cheap, you could run on a single P40 GPU with AQLM quantization, but this would be quite slow (~1 tok/s).
They have no money
Interesting that no-one is factoring in the electricity costs for all these self-host solutions. Guess that’s one consolidated invoice which is not traceable to a specific machine anyway…
You just touched on a subject that a friend who works for a utility company mentioned the other day as fallout from Remote work. AC/Heating units are now running round the clock 'cos folks have to use them non stop. The result has been astronomical utility bills for many.
Note, Discourse hosted plans get related topics included included in plan. This is powered by AI embeddings.
Some providers such as Cloudflare do provide a free tier with a generous amount of usage for self hosters.