SparseGPT can cut model memory and inference compute roughly in half for massive GPT models, enabling cheaper hosting and faster inference without retraining. Joint sparsity+quantization can match lower-bit storage with better accuracy than pure quantization.

SqueezeLLM cuts model storage and single‑request latency by ~2× while keeping near‑FP16 quality, enabling cheaper and faster on‑prem or cloud inference for generative LLMs.

HomoDistil produces smaller, better-performing BERT derivatives by pruning from the teacher while distilling; this saves storage and lowers fine-tuning costs while preserving quality—useful when you need compact models with higher accuracy than typical distilled alternatives.

You can cut 30–50% of many pre-trained transformers with one cheap pruning pass and little accuracy loss, saving memory and inference costs without costly retraining.

Bonsai makes structured LLM compression feasible on commodity GPUs, cutting memory needs and producing faster models so teams can reduce inference cost and enable on-device fine-tuning without enterprise hardware.

LoRAPrune cuts pruning memory and gives real GPU latency wins while keeping better accuracy than prior structured-pruning methods, enabling practical deployment of much larger LLMs on fewer GPUs.

Sparsity plus SquareHead can reduce LLM inference latency and cost on CPUs/GPUs (2–8x) while keeping accuracy for many tasks, enabling cheaper deployment on commodity hardware.

E-Sparse cuts LLM GPU memory by ~43% and speeds matrix work 1.24–1.53× on Ampere hardware, letting teams host larger models or reduce instance costs with small accuracy trade-offs.

PAFT can preserve both task accuracy and alignment without retraining large models end-to-end; companies can run SFT and alignment in parallel, sparsify adapters, and merge them to ship stronger, aligned models faster.

BESA makes aggressive pruning of large LLMs practical on a single A100 GPU, preserving model quality and enabling lower-cost deployment or faster inference when paired with quantization.

If you compress models with pruning and block-wise quantization, order and method choice change accuracy and thus service quality; using sparsity before quantization (S → Q) is an easy rule to reduce avoidable accuracy loss.

TEAL cuts memory movement during single-batch decoding and delivers up to ~1.8× throughput gain without retraining, lowering inference cost for edge or low-latency deployments.

ComPEFT slashes adapter size and transfer time so you can host many more task experts per GPU, reduce bandwidth costs, and cut serving latency without retraining.

OWL lets you prune large language models up to ~70% while keeping useful quality and delivering real CPU speedups, enabling cheaper, faster inference and easier deployment on constrained hardware.

Quest reduces memory bandwidth and decode latency for very long-context LLM calls, lowering GPU cost per request and improving responsiveness for document-heavy applications.

Flash-LLM cuts inference GPU cost and increases throughput for production LLM serving by enabling practical unstructured sparsity on tensor cores.

If you can deploy on event-driven or neuromorphic hardware, combining sparse activations with weight pruning can cut inference work dramatically without large quality loss, lowering energy and latency for low-power or real-time apps.

DSFormer reduces transformer model size substantially (2x–3.6x) while keeping accuracy close to original models and can be stacked with distillation/quantization to cut hosting or edge deployment costs further.

SpIEL lets teams fine-tune large LLMs with much less extra GPU memory by tuning only a sparse set of parameters, enabling on-prem or single-GPU adaptation and cheaper experimentation under quantization.

Using a context-aware window k = α·n^C can cut attention compute while keeping accuracy; dynamic windows can be tuned to a given compute budget and often beat tiny fixed windows on long inputs.

This method makes extreme quantization and M:N sparsity reliable, letting you cut model storage and arithmetic cost while preserving accuracy—so you can deploy larger models under tight memory/energy budgets.

Pruning cuts model size and theoretical compute but doesn't guarantee runtime speedups; measuring on your hardware and considering smaller architectures first saves cost and deployment time.

GQSA delivers multi× inference speedups and big memory savings for 7B–30B class LLMs while often preserving or improving zero-shot accuracy, enabling cheaper, faster serving on GPUs and more viable edge deployment.

PGB cuts model compression time from days to hours while keeping most task accuracy, so teams can produce and deploy smaller BERT models faster and with lower compute cost.