Open weights got close enough to run at home this year, not by needing more RAM but the reverse: sparse attention, MoE, latent KV compression, multi-token prediction and four-bit quant. submitted by /u/mattjcoles [link] [comments]