You can also run Mixtral, at a decent token rate, on a post-2020 Apple Macbook Pro M1/M2/M3 with 32GB+ of RAM. 16GB RAM also works, sort of ok, which I suspect is the same quantization a 3090 is using, but I do notice a difference in the quantization. On my M2 Pro, the token rate and intelligence feels like GPT-3.5turbo. This is the first model I've started actually using (vs playing around with for the love of the tech) instead of GPT-3.5.

An Apple M2 Pro with 32GB of RAM is in the same price range as a gaming PC with a 3090, but its another example of normal people with moderately high performance systems "accidentally" being able to run a GPT-3.5 comparable model.

If you have an Apple meeting these specs and want to play around, LLM Studio is open source and has made it really easy to get started: https://lmstudio.ai/

I hope to see a LOT more hobby hacking as a result of Mixtral and successors.