aimet vs llm-compressor
Neural network compression tools compared
aimet and llm-compressor are Python toolkits that compress trained models to reduce memory use and improve inference efficiency. aimet targets PyTorch and ONNX workflows across neural networks, while llm-compressor focuses on preparing Hugging Face language and multimodal models for vLLM.
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.
llm-compressor: Compress Models for vLLM Inference
LLM Compressor applies post-training quantization and related model transformations to prepare Hugging Face models for vLLM deployment. It suits teams seeking smaller or more inference-efficient checkpoints, with support for multiple formats and large-model workflows.
| aimet | llm-compressor | |
|---|---|---|
| Language | Python | Python |
| License | — | Apache-2.0 |
| Stars | 2.7k | 3.8k |
| Forks | 468 | 675 |
| Last analyzed | Oct 14, 2026 | Oct 3, 2026 |
Key differences
- aimet supports PyTorch and ONNX models, while llm-compressor focuses on Hugging Face models prepared for vLLM.
- aimet combines quantization with methods such as spatial SVD and channel pruning; llm-compressor offers weight, activation, attention, and KV-cache quantization.
- llm-compressor lists formats including INT4, INT8, FP8, NVFP4, and MXFP4; aimet describes techniques including calibration, AdaRound, SeqMSE, and Cross Layer Equalization.
- llm-compressor saves checkpoints in `compressed-tensors` format and includes distributed processing and disk offloading options; aimet provides APIs and visualization tools for PyTorch pipelines.
- aimet lists its license as NOASSERTION, while llm-compressor lists Apache-2.0.
- aimet has 2.7k stars and 466 forks, while llm-compressor has 3.8k stars and 675 forks; both are not archived.
Choose aimet if you…
- need to optimize PyTorch or ONNX neural networks beyond language models.
- want compression options such as spatial SVD, channel pruning, or per-layer compression-ratio selection.
- need quantization-aware training or tools to inspect layer sensitivity and weight ranges.
Choose llm-compressor if you…
- prepare Hugging Face models for deployment with vLLM.
- need formats or methods for quantizing weights, activations, attention, or KV caches.
- work with large models and want distributed processing, disk offloading, or architecture-specific examples for MoE and multimodal models.
Compare the numbers side by side, or add more projects
This comparison is generated with AI from the OSRepos analyses of both projects. Always check each project's repository and documentation before choosing.