Home / Docs / Computer Vision / Quantized Vision Encoders

Quantized Vision Encoders (QVE)

High-efficiency, sub-8-bit computer vision models optimized for edge inference, real-time object detection, and embedded deployment.

Stable PyTorch / JAX / ONNX MIT License Last updated: Dec 14, 2025

Overview

Quantized Vision Encoders (QVE) represent Aevum Zenth's next-generation approach to efficient visual feature extraction. By combining hardware-aware mixed-precision quantization, knowledge distillation, and sparse attention mechanisms, QVE models achieve near-FP16 accuracy at INT4/INT8 compute footprints.

Designed for edge devices, drones, autonomous systems, and bandwidth-constrained environments, QVE eliminates the traditional accuracy-efficiency tradeoff without requiring per-device calibration.

⚡ Sub-Millisecond Inference

Optimized kernel fusion and memory-aware tiling enable <1.2ms latency on modern NPUs and mobile GPUs.

📦 85% Model Size Reduction

Dynamic quantization with structured sparsity compresses weights from 2.1GB to ~280MB with <1.2% mAP degradation.

🔒 Zero-Calibration Deployment

Self-stabilizing activation ranges eliminate the need for dataset-specific calibration passes.

Architecture

QVE builds upon a modified Swin-V2 backbone with three key innovations:

  • Layer-Adaptive Quantization - Early layers operate at INT8, while transformer blocks use INT4/INT6 hybrid schemes based on activation variance.
  • Gradient-Aware Weight Clustering - Post-training quantization that preserves high-sensitivity channels via differentiable clustering.
  • Hardware-Router - Automatic kernel selection based on target compute (NPU, DSP, GPU, CPU).

📐 Mathematical Foundation

The quantization function uses a non-uniform affine mapping: Q(x) = round((x - min_s) / Δ) + ZP, where step size Δ is dynamically adjusted per feature map channel during QAT (Quantization-Aware Training).

Performance Benchmarks

Evaluated on COCO 2017 val set and ImageNet-1K. Latency measured on Jetson Orin NX (INT8 inference, TensorRT 8.9).

Model Params (M) mAP (COCO) Top-1 Acc (IN1K) Latency (ms) Size (MB)
QVE-Tiny 4.2 44.1 78.3 0.8 18
QVE-Small 11.8 51.2 82.7 1.1 46
QVE-Base 28.4 56.8 85.9 1.9 112
Swin-T (FP16 Baseline) 28.3 57.2 87.1 4.3 556

Installation & Quick Start

QVE is available via pip, Docker, and edge SDKs. Supports Python 3.9+ and CUDA 11.8+.

bash 
pip install aevum-qve
curl -L https://models.aevumzenth.com/qve-base-int8.pt -o qve-base.pt
python 
from aevum_qve import QVEEncoder, QuantizeConfig

# Load quantized model with hardware-aware routing
config = QuantizeConfig(precision="int8", backend="tensorrt")
model = QVEEncoder.load("qve-base.pt", config=config)

# Inference
image = load_image("sample.jpg")
features, embeddings = model.encode(image, output_layers=[3, 5])
print(f"Latency: {model.benchmark(image, runs=100):.2f}ms")

API Reference

QVEEncoder.encode()

Core inference method. Accepts tensor inputs of shape [B, C, H, W] and returns hierarchical feature maps.

  • image (Tensor) - Input image or batch
  • output_layers (List[int]) - Stages to return (default: [3])
  • return_embeddings (bool) - Whether to compute global CLS token (default: False)

Hardware Support Matrix

Platform Framework Quantization Status
NVIDIA Jetson TensorRT INT8 / FP16 Production
Apple Silicon CoreML / MPS INT8 Production
Qualcomm AI SNPE / QNN INT8 / INT4 Preview
Generic CPU OpenVINO / ONNX INT8 Production