Retour au blog
Technique 9 min

Quantization and Distillation: Smaller, Faster, Cheaper AI

Do you really need a 175-billion-parameter model to classify an email or pull a date from an invoice? In 2026, the answer is no. Three families of techniques (quantization, distillation, pruning) shrink a language model by a factor of 4 to 10, cut its inference cost by as much, and let it run on an ordinary server, a PC, sometimes a phone. This is the shift from demo AI to production AI, the kind that has to hold a budget and a latency.

Article generated by AI. Content written with the help of an artificial intelligence model and reviewed by a human before publication. The figures cited point to their sources, listed at the end of the article.

Model Compression by the Numbers

−75%
Less memory with 4-bit quantization (under 1% loss)
97%
Of BERT's performance kept (model 40% smaller)
10 to 30x
Cheaper to serve a small model than a large LLM (NVIDIA)

Why Compress

A big general-purpose model impresses in a demo, but in production it costs a lot: scarce GPUs, high memory, latency, an API bill that climbs with volume. Most enterprise tasks (classify, extract, summarize, route, answer from a document base) use only a fraction of that power. The goal is not the biggest model, but the smallest model that does the job.

Three main levers cut size and cost without starting from scratch: quantization (lower the precision of the numbers), distillation (transfer knowledge from a large model to a small one) and pruning (remove what is not needed). You can combine them on the same model.

The concrete gain

Quantization is the fastest lever. Moving a model from 16 bits (FP16) to 4 bits (INT4) divides its memory footprint by four, that is 75% less memory, with an accuracy loss often under 1% when using methods like GPTQ or AWQ. A 7-billion-parameter model goes from about 14 GB to under 5 GB, and fits on a consumer card.

The Techniques, Explained

Four approaches dominate, plus one inference trick. Most are available as open tools, which puts them within reach of a modest team.

Quantization

Lower the precision of the weights: from 32 or 16 bits to 8, 4 or fewer. INT4 divides memory by 4. The loss stays small with GPTQ and AWQ.

75% less memory

Distillation

A large model (the teacher) trains a small model (the student) to reproduce its answers. DistilBERT: 40% smaller, 60% faster.

97% performance kept

Pruning

Remove the weights, neurons or attention heads that contribute little. The model slims down without heavy retraining.

lighter network

Purpose-built SLMs

Models designed small from the start (Phi, Gemma, Llama 3.2 1B/3B), often better than a large model on a targeted task.

1 to 8B params

Speculative decoding

A small model proposes, a large model verifies. Generation speeds up without changing the final quality.

faster generation

Quantization: The Most Immediate Lever

Quantization lowers the number of bits used to store each weight of the model. A 16-bit floating-point weight becomes an 8 or 4-bit integer. Memory drops proportionally: INT8 halves it, INT4 quarters it. Less memory means fewer GPUs, models that fit on modest hardware, and faster inference.

The obvious question is the quality loss. Two methods make it negligible. GPTQ (Frantar et al., ICLR 2023) quantizes layer by layer by minimizing the error, to the point of running a 175-billion-parameter model in 3 bits on a single GPU. AWQ (Lin et al., MLSys 2024, best paper award) starts from an observation: about 1% of the weights carry most of the performance. By protecting them, 4-bit quantization becomes nearly lossless.

Distillation: Passing On the Know-How

Distillation trains a small model (the student) to imitate a large model (the teacher), not on raw labels but on the teacher's probability distributions, which are richer in information. The founding case, DistilBERT (Hugging Face, 2019), keeps 97% of BERT's understanding with 40% fewer parameters (66 million versus 110) and a 60% faster inference.

The technique took a leap in 2025 with DeepSeek-R1. Its authors distilled the reasoning abilities of the large model into Qwen and Llama models from 1.5 to 70 billion parameters, starting from about 800,000 verified reasoning trajectories. The result: the 32-billion distilled version rivals much heavier production models on math and code benchmarks.

Keep in mind

Compressing is not free. Quantization that is too aggressive degrades complex reasoning and math tasks faster than classification or summarization. A distilled model inherits the limits and biases of its teacher. The rule: pick the compression level according to the task, and measure quality on your own data before going to production.

What It Changes for the Business

The economic case is direct. According to NVIDIA, serving a small 7-billion-parameter model costs 10 to 30 times less than a 70 to 175-billion model, needs fewer GPUs and specializes in a few hours rather than weeks. At a high call volume (an agent handling thousands of requests a day), the bill gap becomes decisive.

Less hardware

A 4-bit model fits on a consumer card or a modest server. No more reserving scarce, costly GPUs for simple tasks.

A controlled bill

10 to 30 times cheaper to serve is the difference between a pilot that stays a pilot and a system deployed at scale.

The market is following this shift. Valued at $7.76 billion in 2023, the small language model market is expected to reach $20.7 billion in 2030, a 15.1% annual growth rate (Grand View Research).

Use Cases

Document extraction

Reading invoices, contracts, forms: a small quantized model is enough, and runs on premise without sending the data out.

on premise

Customer service

Classify, route and answer common high-volume requests, without an API bill that explodes with traffic.

high volume

Embedded AI

Compressed models on a PC, industrial gateway or phone, with no permanent cloud connection.

offline

Specialized agents

An agent that calls tools does not need a giant model: a fast SLM does the essential work.

10 to 30x cheaper

Belgium & Europe: Sovereignty and Cost

For a Belgian SME or manufacturer, model compression hits two nerves. First, cost: a quantized model runs on affordable hardware, which makes AI accessible without an unlimited cloud budget. Second, sovereignty: a small model deployed on premise processes the data in house, without sending it outside the European Union, in line with the data-minimization logic of the GDPR.

Governance still applies. A compressed model that feeds decisions must be documented and monitored: under the EU AI Act, keep a record of the data, watch for drift and keep a human in the loop on sensitive calls. A smaller model, better controlled and hosted on your side, is often simpler to govern than a large remote one.

Data sovereignty

A light model runs on premise: the data stays in house and never leaves the EU.

Financial accessibility

Compression brings AI within reach of SME budgets, without depending on a variable cloud bill.

Where to Start

1

Start from the task, not the model

Define precisely the work to do (classify, extract, summarize). Most tasks do not require a giant model.

2

Test a small quantized model

Compare a 4-bit SLM (Llama 3.2, Phi, Gemma) against your current solution, on your own data.

3

Measure quality, cost and latency

Quantify accuracy on the task, cost per request and response time, then decide on the right trade-off.

4

Deploy with governance

Move to production on premise or on a sovereign cloud, with drift monitoring and human validation on sensitive cases.

Sources

  1. Sanh et al. (Hugging Face), DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter (40% smaller, 60% faster, 97% of performance kept). arxiv.org/abs/1910.01108
  2. Frantar et al., GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers (ICLR 2023; 175B parameters in 3 bits on a single GPU). arxiv.org/abs/2210.17323
  3. Lin et al. (MIT), AWQ: Activation-aware Weight Quantization (MLSys 2024, best paper award; ~1% critical weights, 4-bit nearly lossless). arxiv.org/abs/2306.00978
  4. Belcak et al. (NVIDIA Research), Small Language Models are the Future of Agentic AI (a 7B SLM is 10 to 30x cheaper to serve than a 70 to 175B model). arxiv.org/abs/2506.02153
  5. DeepSeek-AI, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs (distillation into Qwen and Llama from 1.5 to 70B, ~800,000 reasoning trajectories). arxiv.org/abs/2501.12948
  6. Grand View Research, Small Language Model Market Report ($7.76B in 2023, $20.7B in 2030, 15.1% CAGR). grandviewresearch.com

What is the smallest model that would do your job?

Molderez Consult helps Belgian companies choose, compress and deploy efficient AI models, on premise and at a controlled cost.

Discuss my use case