They are. Working on a product using a fine-tuned Mistral-7B-Instruct-v0.2 model and it's pretty mind-blowing. Works flawlessly on my RTX3090 and serviceable on my M1 MBP as well. I'm building in Rust (using the candle crate), but for personal usage Python is probably the better choice since it's easier to get up and running.
It works great, until recently WizardLM 2 came out and I tried it with local RAG. The difference is quite significant: https://twitter.com/chxy/status/1780101542311579942
Considering switching the default model after they put it back on HF.
I have infra in my house, not gonna lie it cost a lot, I have a rack with 30k of equipment in it (including 5 GPUs)
But this would probably run on an AWS P2 instance with is 0.90 USD an hour, or there’s lambda labs which is also pretty cheap (no affiliation, just satisfied customer)
The most expensive piece in either build is a video card. You want to be able to load the LLM file into your video card RAM.
I just got a 16gb new card for this. I can load up to 34b models on it but poorly. Anything 13b or less runs perfectly. A 12gb card would be able to run 7b models and with the right training I think 7b models can be awesome.
New 4060ti Used 3090