# intel/neural-compressor

SOTA low-bit LLM quantization (INT8/FP8/MXFP8/INT4/MXFP4/NVFP4) & sparsity; leading model compression techniques on PyTorch, TensorFlow, and ONNX Runtime

Repository: https://github.com/intel/neural-compressor
Canonical: https://ross.abutalabs.com/products/neural-compressor
Homepage: https://intel.github.io/neural-compressor/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: low-precision, pruning, sparsity, auto-tuning, knowledge-distillation, quantization, quantization-aware-training, post-training-quantization, smoothquant, large-language-models, awq, fp4, gptq, int4, int8, mxformat, sparsegpt
Last push: 2026-08-26T08:40:37+00:00

## Health v2 (maintenance only)
Score: 92/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 79, longevity 100
- inputs: {"age_days": 2234, "days_push": 7, "days_rel": 63, "gap_med": 64.5, "n_releases_24m": 9}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2704, forks 320 (observed 2026-08-28T04:07:11.573877+00:00)

## What it is
Intel Neural Compressor is an open-source Python library providing state-of-the-art low-bit quantization (INT8/FP8/MXFP8/INT4/MXFP4/NVFP4), sparsity, and other model compression techniques for deep learning frameworks including PyTorch, TensorFlow, JAX, and ONNX Runtime. It supports advanced quantization of large language models and vision-language models with extensive optimization for Intel hardware (Gaudi, Xeon, Core Ultra, Data Center GPUs) and limited support for AMD, ARM, and NVIDIA hardware.

## Use cases
- quantize an LLM to INT4 or FP8 for faster inference
- apply SmoothQuant or GPTQ to compress a large language model
- run post-training quantization on a PyTorch model
- prune or sparsify a deep learning model
- optimize models for Intel Xeon CPUs or Gaudi accelerators
- quantize models to MXFP4 or NVFP4 low-precision formats
- compress a vision-language model like LLaMA or Qwen
- perform quantization-aware training on TensorFlow models

## When to choose
- you need low-bit quantization of LLMs or VLMs across multiple data types
- you target Intel hardware (Gaudi, Xeon, Core Ultra, Data Center GPUs) for inference optimization
- you work across PyTorch, TensorFlow, JAX, or ONNX Runtime and want a unified compression API
- you need sparsity, pruning, or knowledge distillation alongside quantization

## When to avoid
- you only target NVIDIA GPUs and prefer ecosystem-native tools like TensorRT-LLM
- you need a simple one-line quantization without tuning and prefer lighter-weight tools like AutoRound directly
- your framework is outside PyTorch, TensorFlow, JAX, or ONNX Runtime

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, llm-inference, llm-training
- domain: machine-learning, deep-learning, large-language-models, developer-tools
- platform: python, cross-platform
- tags: quantization, model-compression, sparsity, pruning, knowledge-distillation, post-training-quantization, quantization-aware-training, smoothquant, gptq, awq, int8, int4, fp8, mxfp4, pytorch, tensorflow, onnx-runtime, intel-hardware, linux, gpu

## Member repositories
- intel/neural-compressor (main) score 92

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:11.573877+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:15:54.095366+00:00, confidence not recorded.
  - readme: https://github.com/intel/neural-compressor (fetched 2026-08-28T04:07:11.573877+00:00, sha a586ca19d722)
  - homepage: https://intel.github.io/neural-compressor/ (fetched 2026-08-29T09:58:53.987377+00:00, sha 44136fa355b3)
  - registry_pypi: https://pypi.org/pypi/neural-compressor/json (fetched 2026-08-29T09:58:53.996188+00:00, sha 2dfca99a671d)
- Data as of 2026-08-30T08:39:29.467469+00:00.