Show HN: Spectra – A drop-in optimizer that cuts LLM training VRAM by 50%

Spectra Labs introduces SpectraAdamW, a memory-efficient optimizer for LLM fine-tuning that reduces optimizer state VRAM overhead by 50% using factorized variance approximations and frequency-domain momentum compression. The approach maintains full-precision convergence while offering a drop-in PyTorch replacement for standard AdamW, enabling fine-tuning of large models on consumer hardware with significantly lower memory requirements.