[cma]

It beats llama on the benchmark posted below (though maybe leaked into training data). But also you can run it on cheaper split up hardware with less individual vram than the big llama.