Samsung Open-Sources LittleBit LLM Compression Method
Samsung has open-sourced an AI compression method named LittleBit that shrinks a 13-billion parameter large language model into less than one gigabyte. The technique uses latent factorization to compress weights below one bit each and replaces heavy floating-point multiplications with simple bitwise XOR sign flips. According to the material, this approach achieves up to 11.6x faster inference than standard FP16 models while reducing memory and bandwidth requirements. Open article →