AI Infrastructure

LLM Quantization for On-Premises AI: Cut GPU Cost Without Quietly Losing Accuracy

Quantization decides how many GPUs an on-premises AI platform needs. A practical guide to FP8, INT8 and 4-bit weight-only formats, and the governance around them.

Most on-premises AI budgets are decided by one number: how much GPU memory the platform needs. That number is not set by the model you chose. It is set by the numeric format you serve it in.

The same open-weight model can occupy roughly two bytes per parameter in 16-bit precision, about one byte in an 8-bit format, and around half a byte in a 4-bit weight-only format. On a fixed hardware footprint, that difference decides whether a department gets one model or four, whether a second model can be held resident for failover, and how many concurrent sessions a node can serve before requests start queuing.

Quantization is therefore an infrastructure procurement decision that is usually made by an engineer editing a launch flag. This is a guide to making it deliberately.

What the numeric format actually changes

Three things move together when precision drops.

Memory footprint. Weights shrink roughly in proportion to bits per parameter. This is the effect that shows up in the hardware quote.

Memory bandwidth pressure. Decoding a token requires reading the weights. Smaller weights mean fewer bytes read per token, which is why low-precision formats often improve latency on bandwidth-bound single-stream generation.

Available key-value cache. Whatever memory the weights do not occupy is available for the KV cache that holds active sessions. In practice this is the effect that matters most for a shared enterprise platform, because concurrency is what determines whether one node serves a business unit or three.

The trade is accuracy, and it is not uniform. A format that is invisible on summarisation can be visible on numeric extraction from a scanned invoice, or on a long tool-calling chain where one malformed argument ends the run.

The formats worth evaluating

FormatTypical useWhat to watch
BF16 / FP16Reference baseline for evaluationHighest memory cost per model
FP8 (weights and activations)Default on accelerators with native FP8 supportRequires recent GPU generations
INT8 (weights and activations)Equivalent step on older acceleratorsInteger scaling is more sensitive to outlier activations
4-bit weight-only (AWQ, GPTQ)Fitting larger models on smaller cards; many resident modelsCalibration data quality; task-specific regressions

A large study of the Llama-3.1 family, spanning more than 500,000 evaluations across academic benchmarks and real-world tasks, reported FP8 weight-and-activation quantization as effectively lossless across model sizes, INT8 at roughly 1-3% degradation, and 4-bit weight-only quantization as more competitive than expected. The same work found 4-bit weight-only most cost-efficient for synchronous single-stream setups, while 8-bit weight-and-activation formats performed better under asynchronous continuous batching.

Read that as a prior, not a result. It tells you which experiments are worth running first. It does not tell you how a 4-bit build will handle your policy documents, your German-language claims correspondence, or your JSON schema.

Serving runtimes have made the experiment cheap. vLLM supports FP8, AWQ, GPTQ, compressed-tensors and a long list of other schemes, and FP8 can be produced through several underlying algorithms, so a team can hold a baseline and two candidates and compare them on the same hardware.

Evaluate on your own tasks, not on leaderboards

A quantization decision needs a task-level evidence set before it goes near production traffic. Build it once and reuse it for every model and format change:

  1. A frozen prompt set per workload — retrieval answers, document extraction, classification, tool-calling, and any workload with a strict output schema.
  2. Deterministic scoring where possible. Schema validity, field-level extraction accuracy and tool-argument correctness are countable. Reserve human review for the cases that are not.
  3. Long-context cases. Degradation often appears first at the far end of the context window, not on short prompts.
  4. Non-English cases, if the platform serves them. Multilingual behaviour is not always preserved evenly across formats — the same caution that applies to multilingual private RAG.
  5. A latency and concurrency profile, measured at the concurrency the platform is actually sized for rather than at batch size one.

The comparison that matters is the same prompt set against the same runtime on the same hardware, changing only the numeric format. Anything else compares two variables at once.

Where the memory really goes

Quantizing weights does not remove the rest of the budget. When sizing a node, account for:

  • Weights, in the chosen format;
  • KV cache, which scales with context length and concurrent sessions and is often the binding constraint on a shared platform;
  • Activation and runtime overhead for the serving engine;
  • Embedding and reranking models, if they share the same accelerator — see embedding models and rerankers for private RAG;
  • Failover headroom, so losing one node does not take the service below its committed concurrency.

This is why quantization should be decided alongside, not before, GPU requirement estimation and admission control. A 4-bit build that frees memory for KV cache can raise usable concurrency more than a hardware upgrade would, and it can also mask the fact that the platform has no queueing policy at all.

Treat a quantized build as a governed artifact

The common failure is not accuracy loss. It is losing track of which build produced an output.

A quantized model is a new artifact. It has its own file digest, its own calibration dataset, its own producing toolchain version, and its own evaluation record. If it enters production as a runtime flag rather than a registered version, the audit trail records a model name that no longer identifies what actually ran.

Practical controls:

  • Register the quantized build as a distinct version with its own digest, and carry it through the model artifact supply chain like any other imported weight file.
  • Record the calibration data used to produce it, including its sensitivity classification. Calibration samples are enterprise data.
  • Store the evaluation report against that version, not against the base model, following the same discipline as local LLM versioning and upgrades.
  • Log the serving format in each execution record so a reviewer can reconstruct which numeric build answered a given request.
  • Re-run evaluation on upgrade. A new base model version invalidates the prior quantized comparison.

Under the EU AI Act’s provider and deployer documentation expectations, “which model version handled this decision” is a question that needs an answer from the log, not from an engineer’s memory. Precision format is part of that identity.

How this fits a routed platform

Quantization is most useful when it is not a single global choice. A platform that can route requests across several local models can hold a 4-bit build for high-volume classification and summarisation, and reserve a higher-precision build for extraction and approval-bearing steps, with the routing rule recorded rather than implied.

VDF AI runs local models, private retrieval and agent orchestration inside the customer’s environment. VDF AI Agents call whichever registered local endpoint policy assigns to a step, and VDF AI Networks records the model, tool and fallback decisions taken across a workflow — so the serving build behind each step stays visible after the fact.

The engineering rule is simple: quantize to buy concurrency, prove it on your own tasks, and register the result as a version rather than a flag.

Sources and further reading


Sizing GPU capacity for an on-premises AI platform? Book a VDF AI architecture review to set precision targets per workload, build the evaluation set, and plan capacity against real concurrency.

Frequently asked questions

What is LLM quantization in plain terms?

Quantization stores a model's weights, and sometimes its activations, in a lower-precision numeric format than the one it was trained in. A model held in 16-bit precision needs roughly two bytes per parameter; an 8-bit format needs about one, and a 4-bit weight format about half. Less memory per parameter means a given model fits on fewer or smaller GPUs, and more of the GPU's memory bandwidth is available for serving concurrent requests.

Does quantization reduce answer quality?

It depends on the format and the workload. A large evaluation of the Llama-3.1 family across more than 500,000 evaluations reported FP8 weight-and-activation quantization as effectively lossless, INT8 in the range of 1-3% accuracy degradation, and 4-bit weight-only quantization more competitive than commonly assumed. Those are aggregate academic results, not a guarantee for your prompts. Structured extraction, long-context retrieval and tool-calling workloads deserve their own evaluation before promotion.

Which quantization format should an enterprise start with?

On GPUs with native 8-bit floating point support, FP8 is the usual starting point because it halves memory against 16-bit with minimal reported quality change. On older accelerators without FP8 support, INT8 is the equivalent step. Four-bit weight-only formats such as AWQ or GPTQ are worth using when a model must fit on a smaller card or when you need many models resident at once, and they should be evaluated per task rather than adopted platform-wide by default.

How does quantization affect on-premises capacity planning?

Model weights are only part of the memory budget. Key-value cache for concurrent sessions, activation memory, the serving runtime and headroom for failover all consume capacity. Quantization lowers the weight component and frees memory for KV cache, which usually raises the number of concurrent sessions a node can hold. The right way to size is to fix a target concurrency and latency, then test which format meets it, rather than sizing from parameter count alone.

Does a quantized model need separate governance?

Yes. A quantized build is a distinct artifact with its own digest, calibration data lineage and evaluation record. Treat it as a separate registered version rather than a configuration flag, so audit records show exactly which numeric format produced a given output and evaluation evidence can be tied to the build that is actually serving traffic.

Filed under
local LLMquantizationGPU capacityon-premises AImodel governanceAI infrastructure
On-Prem AI

Plan your on-prem AI deployment

Book an architecture call and we will scope a private, on-prem AI deployment for your environment — integrations, hardware, and governance included.

View the deployment roadmap

Keep reading