Model compression makes an artificial intelligence model smaller or cheaper to run. Common approaches include quantization, pruning, knowledge distillation and architectural changes that reduce the amount of computation or memory needed for inference.
Compression can make capable models practical on personal devices and lower-cost servers, but it may reduce accuracy, reasoning quality or reliability on some tasks. Good evaluation compares the compressed model with the original across representative tasks rather than assuming that a smaller file preserves every capability.
Acronyms and aliases
AI model compression synonymartificial intelligence model compression variant
Related terms
Frequently asked questions
How can an AI model be compressed?
Developers can use methods such as lower-precision quantization, pruning less useful weights, distilling knowledge into a smaller model or designing a more efficient architecture.
Does model compression always reduce quality?
Not always in a noticeable way, but some information or numerical precision is usually removed, so quality must be measured on the tasks and hardware that matter.
