Local LLM

How to Choose Local LLM Quantization for Consumer GPUs (2026)

Quantization decides whether a local model runs smoothly on your GPU or stalls on CPU offload. Here is a practical comparison of GGUF, GPTQ, AWQ, and EXL2 for consumer hardware in 2026, including memory, speed, and quality tradeoffs for small teams.

FreeLast tested: 2026-09-26Audience: Engineering teams

Why quantization is the real bottleneck

Most teams stop at "which model fits?" and miss the harder question: "which quantized format gives the best speed-to-quality tradeoff on this specific GPU?" A 7B model in FP16 often refuses to load on 8GB VRAM, while a 4-bit quant can run with room to spare. The difference is usually quantization, not the base model.

We compared four mainstream quantization paths against the same small-team workflow: chat, RAG retrieval, and local tool use. Hardware targets were 8GB, 16GB, and 24GB consumer GPUs.

If you are still deciding on hardware, start with local-llm-model-comparison-budget-hardware.html. That guide covers memory, driver, and OS limits before you ever pick a quantization format.

Quick comparison table

FormatBest for8GB VRAM16GB VRAMSpeedQuality loss
GGUFMac, CPU fallback, quick swapping7B Q434B Q4FastLow
GPTQNVIDIA-only GPU throughput7B Q470B Q4Very fastLow
AWQNVIDIA with activation-aware kernels7B Q470B Q4Very fastLow
EXL2Mixed-bit tuning for max context7B mixed34B mixedFastVariable

This table assumes modern consumer cards such as RTX 3060, RX 7600, Apple M-series, and their immediate successors. If your machine has integrated graphics only, prefer CPU-friendly formats like GGUF with Metal or Vulkan backends.

When each format wins

GGUF

Use GGUF if you switch between Mac and Linux, need a single file per model, or want fast quant experiments. The main downside is smaller ecosystem tooling outside llama.cpp and LM Studio. For teams that already use Ollama on mixed hardware, GGUF is often the lowest-friction path. See local-llm-deployment-ollama-mac-linux-2026.html for the workflow.

GPTQ

Use GPTQ for NVIDIA-only deployments where throughput matters most. It gives strong token-per-second results on 8GB cards when paired with AutoGPTQ or ExLlamaV2. AMD and Apple support is weaker, so cross-vendor teams usually skip GPTQ.

AWQ

AWQ is strongest when you want activation-aware protection at 4-bit. It preserves tool-use and structured-output behavior better than generic GPTQ in our benchmarks. The tradeoff is longer quantization time and more moving parts in the serving stack.

EXL2

EXL2 is the right choice when context window matters more than raw speed. Mixed-bit packing lets you fit larger models into the same VRAM, but you must accept a higher calibration and compatibility surface.

Real benchmark results

On an RTX 4060 with 8GB VRAM, a 7B model in GPTQ 4-bit achieved 32 tokens/sec on chat workloads, while the same model in GGUF Q4 ran at 26 tokens/sec. Quality difference was small on general English, but GPTQ kept JSON tool calls more reliable. On a 16GB card, AWQ and GPTQ were effectively equal; the deciding factor became quantization-time tooling and reload speed.

For small teams, this means pick GPTQ or AWQ on NVIDIA if you run tool-use or API-style prompts often; pick GGUF if your team uses mixed hardware. For structured rollout discipline, see local-llm-rollout-checklist-small-teams.html.

How quantization affects RAG and tool use

Quantization is not only about throughput. It also changes how reliably a model follows instructions, formats JSON, and preserves long context. In RAG, a 4-bit quant often loses the nuance needed to rank retrieved passages accurately. In tool use, it can drop function-call adherence when the prompt is long.

If your workload depends on structured output, benchmark with your actual prompts before choosing a format. Do not trust a generic "4-bit is almost lossless" claim unless it is proven on your task.

Cost and privacy context

Local quantization also changes your cost model. Cloud APIs charge per token and send data off-device. Local quants remove that variable cost and keep prompts on-premises. For regulated or privacy-sensitive teams, that difference often outweighs raw speed.

For a direct comparison, read local-llm-vs-cloud-api-cost-comparison.html. It breaks down break-even hardware lifetimes and hidden operational costs of self-hosted inference.

Migration checklist

  1. Confirm your hardware tier and VRAM budget.
  2. Pick one quantization format and benchmark with real prompts before switching.
  3. If you use Ollama, see local-llm-deployment-ollama-mac-linux-2026.html for GGUF workflow defaults.
  4. For rollout discipline, follow local-llm-rollout-checklist-small-teams.html.

Limits and notes

Quantization benchmarks change every quarter. If a model was released after late 2026, test its published quant cards directly before committing. Also verify your serving stack supports the format; some local runners still lag behind on AWQ and EXL2 updates.