The price of intelligence is dropping to near-zero... An actually-open 72b model that benchmarks near GPT4 (Mistral Medium is still closed source / closed weights, even if an early version of it leaked with miqu) is insane. Even the 14b model easily outperforms gpt-3.5-turbo, and given how useful gpt-3.5-turbo finetunes are (they generally outperform non-finetuned GPT4 at a given task they've been finetuned on), I'd imagine the 14b model will prove pretty useful too as a super-low-cost finetuning target. Not to mention the fact that the 14b and especially 7b models are runnable locally on consumer GPUs...
Wonder if the 0.5B model with int4 quant runs on my router...