Skip to content

Concepts · Quantization

How a datacentre model fits in your office.

Quantization is the unglamorous technology that decides whether sovereign AI costs a server or a datacentre. Here's how it actually works, in enough detail for your engineers to check.

The mechanism

A model's size is bytes per weight × weights.

Training uses 16-bit numbers. Inference doesn't have to — most of that precision is redundancy you can spend down.

Quantization stores each weight in fewer bits — grouped and rescaled so the model's behaviour is preserved. Half the bits, half the memory, and usually faster too: at the low-batch, single-tenant loads on-prem systems actually run, generation speed is limited by how fast weights stream from memory, so a smaller model is a quicker one.

Precision 8B model 70B model Runs on
16-bit (training) ~16 GB ~140 GB Multi-GPU node
8-bit ~8 GB ~70 GB Datacentre GPU(s)
4-bit (the sweet spot) ~4.5 GB ~40 GB One workstation GPU / one 48 GB card (tight — or 2×24 GB)
1–2 bit (the frontier) ~1.5–2.5 GB ~12–20 GB Laptop or edge box (a phone, for 27B-class and below)

Weights only — working memory adds the KV cache (below) and runtime overhead. The 70B-on-one-48-GB-card fit assumes a compact 4-bit variant and a quantized cache; exact figures depend on format and grouping, and we size the full budget during scoping.

The trade

What it costs, and how we keep it honest.

Quantization trades a measurable sliver of quality for a 3–4× cut in memory and the hardware tier below it. The operative word is measurable.

4-bit: the production default

Modern 4-bit formats — GGUF K-quants for llama.cpp-style serving, AWQ or GPTQ elsewhere — typically cost a point or two on broad benchmarks; on a narrow, fine-tuned task, often nothing detectable. This is where most sovereign deployments should live: a 70B in ~40 GB of weights, an 8B in under 5 GB.

Below 4 bits: engineered, not naive

Simply rounding harder destroys quality. Purpose-built ternary and binary formats — increasingly paired with quantization-aware training — keep models useful at 1–2 bits per weight, shrinking a 27B-class model to phone size. This frontier is moving fast, and it's the one we watch most closely.

Our rule: re-evaluate, then ship

Every quantized build is re-run against your evaluation set and shipped with the quality delta documented next to the full-precision baseline. If quantization costs more than the deployment gain is worth for your task, the report says so and we size up.

The second budget

Context has a memory bill too.

Weights are only half the sizing exercise. The other half is the KV cache — the model's working memory of your conversation or document.

For every token in context, the model stores key-value activations. At short contexts this is noise; at 100K-token contexts — a long contract, a full case file — the cache grows into the gigabytes and can rival the weights themselves.

The cache quantizes too: 4-bit KV storage cuts its footprint several-fold, which is often the difference between "long documents fit" and "they don't" on a given machine. Architecture matters as well — newer models with hybrid attention keep the cache disproportionately small for their size.

Practical consequence: when we commit to "handles your 200-page contracts on one machine", that claim has been sized — weights, cache, overhead — not hoped.

Why it matters here

Quantization is sovereignty's cost curve.

Every bit shaved off a weight moves "runs where your data lives" one step down the hardware ladder.

At 16-bit, running your own model is a capital request with a floor plan attached. At 4-bit, it's one server. At the 1–2 bit frontier, it starts meaning laptops, edge boxes and machines that have never seen a network cable — the environments where regulated data actually lives.

That's why we treat quantization as a first-class discipline, not a checkbox. See what the resulting systems run on in the sovereign runtime.

FAQ

Quantization questions.

Does quantization make the model worse?
Slightly, and measurably — which is the point. 'Often invisible on a narrow task' isn't a deliverable, so we re-run your evaluation set on the quantized build and ship the delta next to the full-precision baseline. A quantized model without an eval report is a guess.
Why does long context need so much memory?
The model keeps a key-value (KV) cache for every token in the conversation, and at long contexts the cache — not the weights — becomes the biggest memory consumer, growing linearly with context length into the gigabytes. The cache can be quantized too, cutting its footprint several-fold. It's a second budget we size explicitly during scoping, because "supports 100K context" on a spec sheet means nothing without the memory to hold it.

Want the sizing for your use case?

Tell us the documents and the hardware you have — we'll come back with the memory budget and the machine it runs on.