The current generation of models (Llama/Llama2) seem to pass a threshold of "good enough" for the majority of use cases at 60b+ parameters. Quantized 30b model

The current generation of models (Llama/Llama2) seem to pass a threshold of "good enough" for the majority of use cases at 60b+ parameters. Quantized 30b models that can run in 24gb of GPU VRAM are good enough for many applications but definitely show their limitations frequently. It is likely that we will eventually see good fine tunes for Llama 30b that produce usable output for code/other challenging domains, but until we get GPUs with 48g+ VRAM we're going to have to make do with general models that aren't great at anything, and fine tunes that only do one very narrow thing well.