Pushing for LoRA sharing to reduce download waste
Borkato · reddit · 2026-08-27
A Reddit user proposes that creators upload LoRA weights instead of full model files. The discussion highlights the massive waste in storage and bandwidth from downloading full finetunes, arguing that standardizing LoRA distribution on base models would significantly lower resource consumption and improve efficiency.
More from Infra
- QNX partners with Hailo for edge Physical AI: 14x performance consistency — pdamodaran · 2026-08-27
- US Holds 15-20x Compute Advantage, But May Not Matter for Some Threats — ohlennart · 2026-08-27
- Hark partners with NVIDIA for gigawatt-scale compute on Vera Rubin platforms — adcock_brett · 2026-08-27
- Minimax H3 Local Benchmark: 5-Second Clip Takes 4 Minutes on AMD 7900XT — thevictor390 · 2026-08-27
- GLM-5.3-Flash Runs at 160t/s with 1.4M Context on RTX6000 Pro — AutonomousHangOver · 2026-08-27
- llama.cpp adds TENSOR_READ_LAZY, MoE expert tensors no longer need to sit in VRAM — jacek2023 · 2026-08-27