Maple-Preview: A 20B-A1B Ternary-Weight Open-Weight Reasoning LLM
Maple-Preview has been released as a new 20B parameter open-weight large language model (LLM) that utilizes ternary weights for efficient reasoning. The model adopts an "A1B" architecture, indicating it has 1 billion active parameters during inference. This release represents a practical implementation of extreme 1.58-bit style quantization at a 20B parameter scale, making local deployment of reasoning LLMs significantly more hardware-efficient. It demonstrates how ternary weights can reduce memory and compute requirements without sacrificing complex reasoning capabilities. The model combines a large total parameter count of 20B with a highly efficient active parameter count of 1B (A1B) per forward pass. By using ternary weights with values restricted to -1, 0, and 1, it drastically reduces memory bandwidth requirements during inference.
## BACKGROUND
Ternary LLMs, often referred to as 1.58-bit LLMs, restrict weight values to {-1, 0, 1}, replacing resource-intensive multiplication operations with simpler addition and subtraction. The "A1B" designation is associated with Mixture-of-Experts (MoE) architectures where only a fraction of the total parameters (in this case, 1 billion active parameters) are computed per token, further optimizing execution speed and resource usage.