Quick answer: For most AI users in 2026, 16GB VRAM is the minimum you should target. It handles 7B-13B LLMs, Stable Diffusion XL, and basic fine-tuning. If you work with larger models or do frequent training, 24GB is the sweet spot. New to GPU specs? Our GPU VRAM guide for beginners walks through the basics first.
Check NVIDIA GeForce RTX 4070 Ti Super on Amazon→Buy on Shopee SG→VRAM requirements by workload
The amount of VRAM you need depends entirely on what you are running. Here is a comprehensive breakdown:
| Workload | Minimum VRAM | Recommended VRAM | Ideal VRAM |
|---|---|---|---|
| 7B LLM (Q4 quantized) | 6GB | 8GB | 12GB |
| 7B LLM (FP16 full) | 14GB | 16GB | 24GB |
| 13B LLM (Q4 quantized) | 10GB | 12GB | 16GB |
| 13B LLM (FP16 full) | 26GB | 32GB | 48GB |
| 34B LLM (Q4 quantized) | 20GB | 24GB | 32GB |
| 70B LLM (Q4 quantized) | 43GB | 48GB | 80GB |
| Stable Diffusion 1.5 | 4GB | 8GB | 12GB |
| Stable Diffusion XL | 8GB | 12GB | 16GB |
| Flux.1 | 12GB | 16GB | 24GB |
| LoRA fine-tuning (7B) | 12GB | 16GB | 24GB |
| QLoRA fine-tuning (7B) | 6GB | 8GB | 16GB |
| QLoRA fine-tuning (13B) | 10GB | 16GB | 24GB |
| Dreambooth (SDXL) | 12GB | 16GB | 24GB |
How LLM VRAM usage works
A language model’s VRAM footprint follows a simple formula: parameters x bytes per parameter + overhead.
At FP16 (full precision), each parameter uses 2 bytes. A 7B model needs ~14GB just for weights. At Q4 quantization (4-bit), each parameter uses ~0.5 bytes, so a 7B model needs ~3.5GB for weights plus 2-3GB overhead for context and KV cache.
Key insight: quantization is what makes consumer GPUs viable for AI. Without it, only datacenter cards could run meaningful models.
| Quantization Level | Bytes per Parameter | 7B Model | 13B Model | 34B Model |
|---|---|---|---|---|
| FP16 (full) | 2.0 | ~14GB | ~26GB | ~68GB |
| Q8 (8-bit) | 1.0 | ~7GB | ~13GB | ~34GB |
| Q5 (5-bit) | 0.625 | ~5.5GB | ~10GB | ~25GB |
| Q4 (4-bit) | 0.5 | ~4.5GB | ~8.5GB | ~21GB |
| Q2 (2-bit) | 0.25 | ~3GB | ~5GB | ~11GB |
Add 2-4GB overhead for context window, KV cache, and framework memory. Longer context windows (16K+) consume more.
VRAM for image generation
Image generation models have different VRAM characteristics than LLMs:
- Stable Diffusion 1.5: 4GB minimum, but 8GB gives you room for larger resolutions and ControlNet
- Stable Diffusion XL: 8GB minimum, 12-16GB recommended for batch generation and advanced workflows
- Flux.1: 12GB minimum for the full model, 8GB with quantized variants
- ComfyUI workflows with multiple models loaded simultaneously can easily use 16-24GB
- AI video models (HunyuanVideo, CogVideoX, SVD) push VRAM requirements far higher — see how much VRAM for AI video for the full breakdown
Unlike LLMs where you load one model at a time, image generation pipelines often stack multiple models (base model, ControlNet, upscaler, IP-Adapter). Each one adds to VRAM usage.
VRAM for fine-tuning and training
Training consumes far more VRAM than inference because you need to store gradients, optimizer states, and activations in addition to model weights:
- Full fine-tuning uses roughly 4x the model’s inference VRAM — a 7B model needs ~56GB
- LoRA reduces this by only training small adapter layers — a 7B model needs ~18GB
- QLoRA quantizes the base model to 4-bit and trains LoRA adapters — a 7B model needs ~6-8GB
QLoRA is the technique that brought fine-tuning to consumer GPUs. If you plan to fine-tune, check our best GPU for fine-tuning guide for specific hardware recommendations.
GPU recommendations by VRAM tier
8GB — starter tier
- RTX 4060 (~$479): enough for quantized 7B models and Stable Diffusion 1.5
- Suitable for: learning, experimentation, lightweight inference
12GB — capable tier
- RTX 5070 (~$875): GDDR7 bandwidth with Blackwell architecture
- RTX 4070 (~$500): proven Ada Lovelace option
- RTX 3060 12GB (used, ~$250): best budget entry
- Suitable for: quantized 13B models, SDXL, QLoRA on 7B
16GB — recommended tier
- RTX 5070 Ti (~$1,050): best performance at 16GB
- RTX 4070 Ti Super (~$800): best value at 16GB
- RTX 4060 Ti 16GB (~$425): budget 16GB option
- Suitable for: 13B models, SDXL, Flux.1, LoRA fine-tuning, Dreambooth
24GB — enthusiast tier
- RTX 5090 (32GB, ~$4,900+): actually 32GB, the new flagship
- RTX 4090 (24GB, ~$2,200): sweet spot for serious AI work
- RTX 3090 (used, 24GB, ~$820): best VRAM-per-dollar
- Suitable for: 34B models, full LoRA fine-tuning, professional workflows
Which VRAM tier should you target?
- Casual use — chatbots and small image generation? 12GB handles quantized 7B-13B models and SD 1.5/SDXL. The RTX 3060 12GB (used,
$250) or RTX 4060 ($479) gets you started. - Regular LLM inference and Flux image generation? 16GB is the practical minimum. The RTX 4060 Ti 16GB (
$425) is the cheapest new option; the RTX 4070 Ti Super ($800) is faster. - Fine-tuning models or running 34B LLMs? 24GB gives you real headroom. The RTX 4090 (
$2,200) or a used RTX 3090 ($820) are your best options. - Professional workflows or 70B models? You need 48GB+ — either dual GPUs or a workstation card like the RTX 6000 Ada.
Common mistakes when choosing VRAM
- Buying 8GB and planning to “make it work” — CPU offloading is 10-50x slower and defeats the purpose of GPU inference.
- Overlooking quantization as a VRAM strategy — Q4 quantization cuts VRAM usage by 75% with moderate quality loss, making 13B models fit on 12GB cards.
- Not accounting for KV cache and context length — longer context windows (16K+) consume 2-4GB of additional VRAM beyond model weights.
Common mistakes
Buying too little VRAM to save money. A $425 card with 16GB VRAM is more useful for AI than a faster 12GB card at the same money. Always prioritize VRAM over compute speed for AI workloads.
Assuming you can just use system RAM. CPU offloading exists but runs 10-50x slower than GPU inference. It is a fallback, not a strategy.
Ignoring future needs. Models are getting larger and workflows are getting more complex. If 16GB feels tight today, it will feel cramped in a year. Buy one tier above your current minimum when possible.
The bottom line
For AI in 2026, target these minimums:
- Casual inference and experimentation: 12GB
- Regular LLM use and image generation: 16GB
- Fine-tuning and professional workflows: 24GB+
The best GPU for AI under $500 and under $1000 guides can help you find the right card at your budget. For higher tiers, see under $1500 and under $2000. If you’re building a dedicated training rig, our best GPU for AI training at home guide weighs power and thermals.
Check NVIDIA GeForce RTX 4060 Ti 16GB on Amazon→Buy on Shopee SG→