2026 updateThe numbers below are measured on real hardware. For the current 2026 model lineup (Qwen 3, Llama 4, Gemma 3, DeepSeek R1) with an interactive per-quant breakdown, use the VRAM Calculator.
Everything about running local LLMs starts with one number: your GPU’s VRAM. It determines which models you can load, how fast they run, and whether you’ll need to compromise on quality. This guide maps every common VRAM bucket — from 4GB to 80GB+ — to the exact models you can run.
Quick answer: Add ~1.2 GB per billion parameters at Q4_K_M quantization. A 7B model needs ~5 GB VRAM, a 13B needs ~9 GB, 34B needs ~22 GB, and 70B needs ~40 GB at Q4 (or 24 GB at Q2). Always leave 1–2 GB headroom for the KV cache and OS.
The VRAM → model-size matrix
| Your VRAM | Comfortable (Q4_K_M) | Stretch (Q3 / Q2) | Use case |
|---|---|---|---|
| 4 GB (GTX 1650, RTX 3050) | Phi-3 Mini, Qwen 2.5 3B, Llama 3.2 3B | Llama 3.1 7B Q2 (slow) | Basic chat, coding autocomplete |
| 8 GB (RTX 3060 Ti, 4060, 3070) | Llama 3.1 8B, Mistral 7B, Qwen 2.5 7B | Llama 3 13B Q3 | Daily driver chat, coding |
| 12 GB (RTX 3060 12GB, 4070) | Llama 3 13B, Qwen 2.5 14B, CodeLlama 13B | Mixtral 8x7B Q3 | Serious coding, simple agents |
| 16 GB (RTX 4060 Ti 16GB, 4070 Ti Super) | Qwen 2.5 32B Q3, Gemma 2 27B, Mixtral 8x7B Q4 | Llama 3 70B Q2 (painful) | High-quality local assistant |
| 24 GB (RTX 3090, 4090, 7900 XTX) | Qwen 2.5 32B Q5, Llama 3 70B Q2, Mistral Large Q2 | Llama 3.1 70B Q3 | Near-GPT-4 quality at home |
| 48 GB (2× 3090, RTX A6000) | Llama 3.1 70B Q5, Qwen 2.5 72B, DeepSeek-V2.5 | Llama 3 405B Q2 (glacial) | Production-grade local AI |
| 80+ GB (H100, 2× A6000) | Llama 3.1 405B Q4, DeepSeek R1, full-precision 70B | Everything short of frontier | Research, fine-tuning |
Rules of thumb
- Your model must fit in VRAM or speed drops 10-30× when offloading to CPU
- Add 1-2 GB headroom for KV cache (growing context window), framework overhead, and the OS
- Quantization is your friend. Q4_K_M runs at ~94% quality of the original at 30% the size
- VRAM matters more than compute speed for practical local LLM use — an old 3090 (24GB) beats a brand-new 4070 (12GB) for any model over 13B
- Unified memory (Apple Silicon) is a different beast — you can “stretch” beyond traditional VRAM limits but at the cost of speed
Related guides
- 🛠 Interactive hardware checker — punch in your GPU and get instant compatibility
- ⚡ Tokens-per-second benchmarks — real-world speed across GPUs
- 💰 Best GPU for local LLMs — budget-tier recommendations
- 🧮 Quantization explained — what Q4_K_M, Q5, Q8 actually mean
Last updated: 2026-04-22.