Author Predicts Consumer GPUs Will Run Fable-Level Models

andrewchen · x · 2026-07-13

Based on local model experiments, the author boldly predicts that future consumer-grade GPUs might run "Fable-level" models.\n\nThe core arguments are:\n- The relationship between training corpus and model parameters resembles compression; current models already compress vast knowledge into relatively limited parameters.\n- Techniques like quantization, MoE, and pruning continuously improve "capability per parameter." The author cites the so-called Densing Law, where the capability-to-parameter ratio roughly doubles every 3.3 months.\n- If this trend continues, the "irreducible knowledge core" required by models could shrink to tens of GBs, fitting into the VRAM of high-end consumer GPUs.\n\nThe author acknowledges this curve won't hold indefinitely due to compression limits, but notes that today's open-source 27B models are already approaching this threshold.

Original post →

More from AGI Musings

AGI Musings channel →