SoloHacker007/DeepSeek-R1-14B-IndraBit-APoT

Indra-Bit APoT Quantized: DeepSeek-R1-Distill-Llama-70B

This model was converted using the Indra-Bit multiplier-free quantization scheme.

What is Indra-Bit?

Instead of floating-point matrix multiplication, every weight is stored as a 4-term sum of powers of two: sign × (2^k1 + 2^k2 + 2^k3 + 2^k4).

This enables inference using only bit-shifts and addition — zero floating-point multiplications.

Performance

• 3.3x–4.3x speedup over FP32 on CPU (AMD Ryzen 5 5600H, single-threaded)
• ~50% memory reduction (weights stored as float16 APoT)
• 0 floating-point multiply operations during inference

Source

Converted by the Indra-Bit project: https://github.com/karanmertiya/indra_bit_engine