quic

aimet: Quantize and Compress Trained Neural Networks

AIMET is a Python toolkit for quantizing and compressing trained PyTorch and ONNX models. It helps ML teams reduce inference compute and model size while working to preserve accuracy, especially for edge deployment.

View on GitHub Compare
Stars
2.7k
Forks
468
Stars, last 30 days
+20
Last commit
today

Our take

Activeregular commits and releases

Health score

48/100

How it is scored
Activity
81
Community
no data
Issues
44
Pull requests
20

+20 stars in the last 30 days

  • Releases more often than 79% of the projects we track

AIMET is a actively maintained, specialized optimization toolkit for teams already working with PyTorch or ONNX models, with a broad range of quantization and compression techniques.

Good fit if

  • You need to optimize a trained PyTorch or ONNX model for inference size or compute.
  • You want to compare post-training quantization approaches or use PyTorch quantization-aware training.
  • You can work with Python 3.10 or later and want a project that is receiving frequent commits and releases.

Look elsewhere if

  • Your models are outside the supported PyTorch and ONNX frameworks.
  • You need a general-purpose model training or deployment platform rather than model quantization and compression tools.
  • Your workflow depends on a specific technique that AIMET does not support for your framework. For example, OmniQuant is listed as PyTorch-only.
All health signals
Last commit2026-10-13 (today)
Commits, last 90 days100+
Releases, last 12 months28 (latest 2.41.0, 2026-10-05)
Issues closed, last 90 days3 (typically closed in 13 days)
Pull requests merged, last 90 days0
Project age6 years

Checked on 2026-10-14 with the GitHub API.

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

2,750

Forks

479

Change since Nov 17, 2026: 0 stars·0 forks

Overview

AIMET is a toolkit for optimizing trained neural networks through quantization and model compression. It addresses the accuracy and size trade-offs teams face when preparing models for more efficient inference, including deployment on edge devices.

It is intended for ML engineers working with PyTorch or ONNX models who want automated optimization techniques rather than manual layer-by-layer adjustments. Its current development is active, with regular commits and releases.

Key Features

  • Post-training quantization techniques include calibration, AdaRound, SeqMSE, and Cross Layer Equalization.
  • Quantization-aware training is available through the PyTorch package.
  • Supports quantization workflows for both PyTorch and ONNX models, with some techniques limited to PyTorch.
  • Offers compression methods including spatial SVD and channel pruning.
  • Can select per-layer compression ratios automatically.
  • Includes visualization tools for weight ranges and per-layer compression sensitivity.

Use Cases

  • An ML engineer preparing a trained PyTorch or ONNX model for lower-memory or lower-compute inference can evaluate AIMET's quantization and compression methods.
  • A team targeting edge devices can use the toolkit to reduce model footprint before deployment, then assess accuracy against its original model.
  • A model developer facing accuracy loss from basic weight rounding can try advanced post-training quantization methods such as AdaRound or SeqMSE.
  • A PyTorch team able to fine-tune a model can incorporate quantization-aware training into its workflow.

What you need

Detected in the repository

  • Python >=3.10 (from pyproject.toml)
  • Automated checks on GitHub Actions

Getting Started

Install the package for your framework:

pip install aimet-torch

For ONNX, use pip install aimet-onnx. See the README and quick-start guide for setup details.

License in Plain Words

No standard license was detected. Without a license, the default copyright rules apply: you may read the code, but not reuse it. Check the repository or ask the authors.

Considerations

AIMET's Python runtime requirement is 3.10 or later. The provided requirements do not list a mandatory GPU or database, but teams should check the installation guide for environment-specific setup before adoption. The toolkit's scope is model optimization, not a full model-serving or training system. The README identifies the license as BSD 3-Clause; confirm the repository's license terms for your intended use.

Found this useful?

Share it with someone who would like aimet.

Comparisons