hostinglinks.
PRACTICAL TECH, EXPLAINED
Open the LLM VRAM calculator

Local LLM memory requirements

Explore estimated GPU VRAM and system RAM for each model. Compare supported weight formats at an 8K context, or the model's lower native limit. These are planning estimates for one conversation with full GPU offload.

New to running models locally? Start with the Local AI hub. Looking for ChatGPT, Claude or Gemini? Browse current hosted AI models or compare local and hosted AI.

What the catalogue covers

32 model variants, reviewed 14 October 2026. Every model page links to its architecture source. This is a curated selection, not a complete list of open models. Memory figures are planning estimates, not measured performance or a guarantee that a download will run.

Most weight sizes are calculated from approximate parameter counts. gpt-oss and Ministral 3 use rounded published Ollama package sizes. The estimator covers one text conversation and full GPU offload. It does not estimate image processing, concurrent users or multi-GPU behaviour. Gemma 3 and gpt-oss use conservative full-context cache allowances without sliding-window savings.

Models awaiting a suitable calculation

  • Qwen 3.5: hybrid attention needs a separate calculation for its recurrent state.
  • Gemma 4: shared cache and different attention dimensions need architecture-specific estimates.
  • GLM-4.7-Flash: compressed attention cache needs a different formula.

These families are not yet available in the calculator. Adding their parameter counts to the existing formula would not give a reliable estimate. Read the calculation methodology.

Longer context, concurrent users and runtime settings can increase memory use. Check the actual model download before buying hardware.