This page may contain stale information. Last updated: 2026-04-21

Definition

Quantization is a compression technique that reduces the numerical precision of neural network weights and activations. Converting from 32-bit floats to lower precision (8-bit, 4-bit, 3-bit) reduces model size and memory bandwidth requirements.

Key Tradeoff

  • Standard precision (FP32): Full accuracy, high memory/compute cost
  • Quantized (INT8, INT4, INT3): Reduced accuracy risk, but 4-8x smaller models and faster inference

Quantization Approaches

  1. Post-Training Quantization (PTQ): Apply after training (no retraining required)
  2. Quantization-Aware Training (QAT): Include quantization during training
  3. Dynamic Quantization: Different precision for different layers/weights

Challenge: Maintaining Accuracy

Historical problem: Lower precision = accuracy degradation

  • INT8 often acceptable (1-2% accuracy loss)
  • INT4 requires careful calibration
  • INT3 previously considered impractical

Breakthrough: PolarQuant

google’s turboquant introduces novel approach (polarquant) enabling 3-bit quantization without accuracy loss. Specifically targets kv-cache values, where redundancy is high.

Production Applications

  • Model inference (cost reduction)
  • Mobile/edge deployment (size reduction)
  • Long-context LLMs (memory efficiency)
  • Vector search indices (faster retrieval)

Sources