LLM Memory Calculator

This will estimate if a local LLM fits in your GPU VRAM and system RAM

Enter the model, memory and runtime details to check whether it fits:

  • Model Details :
  • Memory and Offload :
  • Runtime :

Live Result :

Verdict :

GPU side total :

System side total :

VRAM available :

Usable unified memory :

Headroom :

Max context at full offload :

Recommended GPU layers :

Offload fits in RAM :

Memory breakdown :

ComponentSizeFormula

Warnings :

    Available as an MCP server for AI assistants

    All sizes are estimates: weights use effective GGUF bits-per-weight averages, the KV cache covers only the full attention layers, the compute buffer is a runtime-dependent heuristic, and batch size scales the compute buffer only. The vocabulary logits buffer holds one fp32 logit per vocabulary entry for each physical batch token, MoE models scale the compute buffer by the active over total parameter ratio, sliding window layers cap their KV cache at min(context, window), and multi-head latent attention models (MLA) are likely overstated because their KV cache is compressed into a low-rank latent vector. Hybrid attention models additionally hold a small fixed state per linear attention layer that is not modeled. Gb values are interpreted as binary GiB (1024^3 bytes).