Can I run this AI model?
Choose your hardware. Everything runs in your browser and nothing is sent anywhere. Results are estimates.
Your hardware
Results
GPU = fits in VRAM (fast). Offload = needs system RAM too (works, much slower; 4 GB is reserved for the OS). No = too big. The number is estimated total GB.
Some links may be affiliate links; we'll say so where they are. How we make money.
Alert me when new models fit my GPU
One short email with the top AI and gaming items we verified that day, plus one useful tool. Signups open soon. Until then, follow along with our RSS feed.
How the estimate works
- Weights = parameters × bits per weight ÷ 8. For example, a 7B model at 16-bit is about 14 GB. Bits per weight are approximate llama.cpp values: Q4_K_M ≈ 4.89, Q5_K_M ≈ 5.70, Q8_0 ≈ 8.5, FP16 = 16 (llama.cpp quantize docs).
- Runtime overhead: 10% of the weights plus 0.5 GB. This is our rule of thumb, not a measurement.
- KV cache (context memory): about 0.0125 GB per billion parameters per 4K tokens. This is a rough average. Real usage depends on layer count, attention type (GQA) and KV-cache quantization, and can be well above or below it.
Parameter counts are the total parameter count Hugging Face reports for each official model repo (linked in the table), fetched Oct 30, 2026. For mixture-of-experts (MoE) models, memory depends on total parameters, not active ones. Some models ship pre-quantized (for example gpt-oss in MXFP4), so their real files can be smaller than the FP16 column. For an exact figure, check the GGUF file size on the model page and add room for context.
Model list updated Oct 30, 2026. Includes 2026 releases (Gemma 4, Qwen3.8, Nemotron 3.5). Very large models (DeepSeek V4, Kimi K3, Muse Spark Open) are left out because they need multi-GPU servers.