← Blog
·2 min read

Quantization and optimization: running open models on smaller hardware

The biggest model isn't always the best choice. Optimization techniques often deliver the same perceived quality at a fraction of the hardware cost.

OptimizationAI architecture

A common mistake when planning an AI project is starting from the largest model available, only to discover that it requires infrastructure disproportionate to the value of the use case. The same perceived quality is often achievable with a smaller, properly optimized model — at a fraction of the cost.

What quantization actually is

A model stores its parameters as numbers with a certain precision (typically 16 bits). Quantization reduces this precision — to 8 bits, or even 4 — proportionally reducing the memory required and often increasing response speed, at a quality cost that for many business tasks is minimal or imperceptible.

The techniques we use most often

How much quality is really lost

It depends on the task, and should be honestly measured case by case: on open-ended conversations or highly complex reasoning tasks, the loss can be noticeable; on more narrowly scoped tasks — extraction, classification, RAG over documents — it's often negligible. That's why, before choosing the quantization level for production, we always test on the client's real use case, not on a generic benchmark.

An illustrative example

In a typical internal-assistant use case over company documentation, moving from an unquantized model to an 8-bit quantized version can cut GPU memory requirements by roughly half, with a quality difference that end-user tests often find indistinguishable. These are indicative figures, not a guarantee: actual sizing should always be verified on the specific project.

Want to talk about it applied to your case?

Book a call →More articles
Keep reading

Open models vs. closed models: what actually changes for your sensitive data

A closed model and an open-weight model aren't two variants of the same product: they change who sees your data and who controls the system. Here's how to decide, case by case.

FADP, GDPR, and artificial intelligence: the compliance checklist for business decision-makers

Adopting an AI tool without checking where the data ends up is the fastest way to turn a productivity gain into a compliance problem. Here's what to check first.