Learn AI /Training and Model Adaptation /Pruning
Pruning
- Deeper
- Training and Model Adaptation
- Technical readers
Also called: model pruning, sparsity.
Back to the milestone you came from
In five seconds
What it is
Pruning removes less useful connections or parameters to reduce model size or runtime while trying to keep nearly the same performance.
In plain words
Pruning trims away parts of a model that seem to matter less, so the model may become smaller or cheaper to run without losing too much quality.
An everyday example
A cook removes ingredients that seem to make little difference, then prepares and tastes the dish again. If the result changes too much, some ingredients go back or the recipe is adjusted. The shopping list becomes shorter only when the removed items are actually left out, and preparation becomes faster only when the kitchen can skip the corresponding steps. Merely crossing ingredients off the page while still handling them saves nothing.
A revision sheet is a human-made summary, while a model’s parameters are not neat, separate facts. Removing one note from a sheet usually has a clear effect, but removing parameters can have indirect effects across many outputs because parameters interact. Also, a crossed-out line on paper always makes the sheet shorter, while a pruned model only becomes faster in practice if the software and hardware can use the sparsity.
Remember thisPruning removes low-importance weights or structures, but zeros alone do not guarantee a smaller or faster model. Retest quality, storage and runtime on the intended deployment system.
Understand it
Understand it
Why it matters
Pruning matters because model size affects cost, speed, memory use and where a model can be deployed. A smaller model may run on cheaper hardware, closer to where it is needed, or with lower latency. But pruning is a trade-off, not a free improvement. If too much is removed, performance can degrade in uneven ways, including on less common languages or edge cases. For managers and practitioners, the key question is not only 'How many parameters were removed?' but 'What happened to quality, reliability and fairness on the tasks we care about, and can our systems actually benefit from sparsity?'
How pruning changes a model
After viewing this, a reader should be able to describe pruning as selecting less important parameters, removing or zeroing them, and then checking whether the model still performs acceptably.
01
1. Start with a dense model
The process begins with a model that has many active parameters or connections.
Draw a rectangle labelled 'trained model' containing a dense grid of dots connected by many grey lines.
02
2. Estimate importance
Each parameter or group is scored using a rule such as magnitude, contribution to loss, or sensitivity to removal.
Overlay small score tags on some lines and dots; show low scores in light grey and high scores in dark blue.
03
3. Remove low-importance parts
Low-scoring parameters are set to zero or removed. The model becomes sparse, meaning many positions are now empty or inactive.
Fade out or strike through selected low-score lines; in the next panel, show many gaps and several labels '0'.
04
4. Test and recover if needed
The pruned model is evaluated. If quality fell too much, a short fine-tuning stage may recover some performance.
Show a small performance meter comparing 'before' and 'after', then a short arrow to a tuning wrench icon labelled 'optional fine-tuning'.
A left-to-right sequence of four panels shows a dense model, an importance scoring step, a pruned sparse model, and evaluation or fine-tuning. Read from left to right to see what goes in, what is removed, and what comes out.
Hover or tap each step to highlight the changed parts and reveal one sentence about trade-offs: smaller model, possible speed gain, possible accuracy loss.
Describe this without the diagram
The diagram is a four-step process. First, a dense trained model is shown with many active connections. Second, the model’s parameters are scored for importance; some are marked as low importance. Third, those low-importance parameters are removed or set to zero, leaving a sparse model with many inactive positions. Fourth, the sparse model is tested, and an optional fine-tuning step can restore some lost performance. The main message is that pruning simplifies a model by removing less useful parts, but the result must be checked because quality can drop.
With reduced motion: A static four-panel diagram shows the same sequence without animation: dense model, importance scores, sparse model after removals, and final evaluation with optional fine-tuning.
Pruning is a way to simplify a trained or partly trained model. The input is usually a model with many parameters, plus a rule for judging which parameters seem less important. A parameter is a numerical setting that affects the model’s output.
What happens next is selective removal. Engineers may set some parameters to zero, remove whole neurons or channels, or remove larger blocks that hardware can skip efficiently. After that, they often test the model again and may fine-tune it so it recovers some of the lost performance.
The output is a sparser model. Sparse means many parameter values are zero or many connections are absent. In some cases this cuts memory use, speeds up inference, or lowers energy use. In other cases, especially if the hardware or software cannot take advantage of sparsity, the practical speed gain is small even though many parameters were removed.
Pruning is often confused with quantisation and distillation. Quantisation keeps the same structure but stores numbers with lower precision, such as fewer bits. Distillation trains a smaller model to imitate a larger one. Pruning usually starts from one existing model and removes parts of it.
How “Shortening a recipe without changing the dish” maps onto the real thing
| In the story | In the AI system |
|---|---|
| ten-page revision sheet | large trained model with many parameters |
| notes that never help on past questions | parameters or connections judged less important |
| crossing out lines | setting parameters to zero or removing parts of the network |
| doing another practice paper | evaluating and possibly fine-tuning after pruning |
A worked example
The regional bureau has an internal model with 12 million parameters that sorts non-sensitive course feedback into eight themes. The training officer and technical focal point test whether magnitude-based pruning will make it practical to run on lower-specification laptops.
Starting from: Original validation accuracy: 92.4%. They propose pruning 40% of individual weights, removing those with the smallest absolute values first.
- They rank the model's weights by absolute value and set the lowest 40% to zero, producing unstructured sparsity.
- They test the pruned model on the same validation set. Accuracy falls from 92.4% to 90.8%.
- They fine-tune the pruned model for two short epochs on the original training set. Accuracy recovers to 91.9%.
- They export the model using a sparse representation, then benchmark memory use and response time on the intended laptops rather than assuming that zeroed weights automatically improve deployment.
Result: The pruned model is accepted after the sparse representation reduces memory use and the quality loss is judged manageable. The benchmark records separately whether response time improves, since the target runtime may not accelerate unstructured sparse weights.
In a United Nations workplace
At a regional bureau town hall, a training officer is asked, without warning, why the course-feedback classifier used for a quarterly learning progress report due at close of business cannot simply be made smaller. The technical focal point proposes pruning the internal model before its next deployment to lower-specification laptops.
Pruning can remove low-importance weights or whole structures, followed by testing and, if needed, brief fine-tuning. This may reduce memory use or runtime while preserving acceptable classification quality. The gain depends on whether the deployment software and hardware can exploit the resulting sparsity.
Before you try thisRaw course feedback containing personal or human resources information must not be uploaded to a tool that has not been approved for that material. Pruning and benchmarking should use an approved environment, with de-identification and access controls where required. Human review remains necessary before themes are reflected in the progress report.
Where you may hear it
- A vendor demo during an ICT governance board meeting
- A deployment review with a regional bureau technical focal point
- A procurement unit's technical evaluation of an on-device tool
- An engineering call after latency and memory testing
Avoid confusion
Avoid confusion
Often confused with
Distillation
Distillation trains a student model to imitate a teacher model’s outputs or other signals; the student is often smaller.
Quantisation
A method for storing and running a model with lower-precision numbers so it uses less memory and often runs faster, with some possible loss in accuracy.
LoRA
A method for adapting a large model by training a small set of added weights instead of changing most of the original model.
Common misconceptions
Common misconception: If 50% of parameters are pruned, the model will run about twice as fast.
Correction: Not necessarily. Unstructured pruning often creates many zeros scattered through the model, but ordinary hardware may still process the same shapes of tensors and gain little speed. Large real-world gains are more common when the runtime supports sparsity well, or when pruning removes whole structures such as channels, heads or blocks.
Common misconception: Pruning and quantisation are basically the same thing.
Correction: They target different parts of the problem. Pruning removes or deactivates parameters or structures. Quantisation keeps them but stores and computes them with lower numerical precision. They can be combined, but one does not automatically imply the other.
Common misconception: A pruned model has forgotten knowledge in a clean, interpretable way.
Correction: Usually not. Parameters interact across many behaviours, so removing some of them does not neatly erase one fact or one skill. Pruning changes the computation distributed across the network, and the effects may appear in unexpected places.
Common misconception: Pruning only happens after training is finished.
Correction: Many workflows prune after training, but not all. Some methods prune during training, gradually encourage sparsity, or alternate between pruning and recovery steps. The term covers several timings and methods, not one fixed procedure.
Quick check
Which statement best describes pruning?
Pruning starts from an existing model and removes or deactivates parameters or structures judged less important. Training a small model to imitate a larger one is distillation. Reducing bit width is quantisation. Adding task-specific capacity is a different adaptation strategy.
Go deeper
Go deeper
The simple version
Pruning trims away parts of a model that seem to matter less, so the model may become smaller or cheaper to run without losing too much quality.
The technical version
Pruning is a family of model compression methods that impose sparsity by removing parameters, activations or larger structures judged less important to the objective. Importance can be estimated by simple heuristics such as weight magnitude, or by criteria related to gradients, Hessian approximations, movement during training, or sensitivity of loss to masking. The pruning target may be unstructured individual weights, structured units such as neurons, channels, attention heads or blocks, or N:M patterns designed for hardware support. After masking or deleting the selected elements, practitioners often perform fine-tuning to recover performance. The practical effect depends on the sparsity pattern and system stack: parameter count may fall, but latency improves only if storage format, kernels and hardware exploit the resulting sparsity. Pruning differs from distillation, which trains a separate smaller student model, and from quantisation, which reduces numerical precision rather than connectivity.
Reviewed external resources
EncyclopaediaPruning (opens in a new tab)Wikipedia
Wikipedia gives a broad overview of pruning concepts and terminology, which suits readers who want a quick orientation before going deeper.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
The Hugging Face LLM Course helps technically minded readers place pruning among other model optimisation methods used in practice.
Deeperlink checked 2026-07-31
CourseShort Courses (opens in a new tab)DeepLearning.AI
DeepLearning.AI short courses are useful for learners who prefer guided, application-focused explanations over reference-style reading.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Pruning (opens in a new tab)YouTube search
The YouTube search results page offers multiple video explanations of pruning, which suits readers who learn well from visual teaching but need to judge quality across sources.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Pruning (opens in a new tab)Reddit search
The Reddit search results page shows community discussions and practical questions about pruning, which suits readers who want to see common confusions and practitioner experiences.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Pruning removes low-importance weights or structures, but zeros alone do not guarantee a smaller or faster model. Retest quality, storage and runtime on the intended deployment system.