SPECTRA compresses any LLM 3x in one click with under 50 MB of RAM

Compress any LLM 3x in one click,near-lossless,in under 50 MB RAM

SPECTRA attacks the memory wall that throttles LLM inference on edge GPUs. Shannon's source-coding bound caps lossless compression at roughly 1.19x for FP16 weights, so hitting 2.8–3.3x requires rate-distortion quantization. SPECTRA combines calibration-free AST saliency scanning, Group-32 4-bit microscaling, kurtosis-driven outlier isolation, and fused Triton W4A16 kernels. It claims near-lossless quality, works on non-transformer architectures like Mamba-2, and runs the whole compression pass on the CPU in under 50 MB of RAM.