Speed, accuracy, and memory across 5 models (Q4_K_M) and a 5-quant sweep of LFM2.5-230M.
Model comparison