Conceptultra-low-bit quantization

Ultra-Low-Bit Quantization

A quantization technique that compresses LLM weights to 1-2 bits, such as ternary weights. It substantially reduces storage, memory bandwidth and matrix multiply-accumulate overhead, but numerical precision is limited and dedicated kernel designs are required.

Timeline (last 90 days)

Oct 21

Ultra-low-bit quantization compresses model weights to 1-2 bits (such as the ternary weights of BitNet b1.58) to reduce storage, memory bandwidth, and matrix multiply-add overhead.

Unverified50%

All Facts (1)

Unverified

Ultra-low-bit quantization compresses model weights to 1-2 bits (such as the ternary weights of BitNet b1.58) to reduce storage, memory bandwidth, and matrix multiply-add overhead.

50%

Source Articles

BitMTP: Co-Accelerating Multi-Token Prediction and Low-Bit LLMsOct 21