AI model training just got 20% cheaper to run, but only if you pick the right layers

A new training method cuts the memory needed to fine-tune large language models by up to 20 percent while keeping accuracy roughly equal to full training. That means researchers and smaller companies can train these models on cheaper hardware instead of buying expensive GPUs.

Large language models require enormous amounts of GPU memory to train, which locks out anyone without serious hardware budgets. This method addresses a real bottleneck, but it works by choosing which layers of the model to update based on gradient analysis — meaning the improvement only kicks in if you pick the right ones. The practical win is narrower than it sounds: you save money and memory, but you need to measure your model's behavior carefully during training to know which layers matter for your specific task.