Conceptternary weights / 三元量化 / 1.58-bit量化

Ternary Weights

An extremely low-precision neural network quantization method that constrains all model weights to the set {-1, 0, +1}. Theoretically requiring only 1.58 bits of storage per parameter, it reduces matrix multiplication to addition and subtraction, substantially lowering storage requirements and improving inference speed. It requires targeted design from the training stage onward.

Timeline (last 90 days)

Sep 16

For ternary weight matrices dominated by zero values, their empirical entropy may well fall in the range of 1.0 to 1.3 bits.

Unverified50%

Aug 5

Ternary quantization means each weight has only three possible value states, typically -1, 0, and +1.

Unverified60%

All Facts (2)

Unverified

Ternary quantization means each weight has only three possible value states, typically -1, 0, and +1.

60%Unverified

For ternary weight matrices dominated by zero values, their empirical entropy may well fall in the range of 1.0 to 1.3 bits.

50%

Source Articles

19.8MB LLM: How a 44M-Parameter Model Achieves 1900 tok/s on CPUSep 15