LOCAL AI / CAPACITY PLANNER
Estimate weights, KV cache, runtime overhead and a practical safety margin before downloading a local LLM.
Minimum estimate: 9.2 GB
HOW IT WORKS
Weights are estimated from parameter count × bytes per quantized parameter. KV cache uses 2 × layers × hidden size × context × concurrent sequences × 2 bytes, assuming FP16 keys and values.
Runtime overhead is estimated at 12% of weight memory with a 1.2 GB floor, then a 20% comfort margin is added. Real use varies by architecture, quantization format, backend, offloading and operating system.
QUICK ANSWERS
No. Apple silicon shares memory between CPU and GPU, but the operating system and other apps also consume it. Leave practical headroom.
Architectures differ in layer count, hidden size, attention design and file overhead. Enter the model-card values for a better estimate.
No. This calculator targets inference. Training and fine-tuning require memory for gradients, optimizer states and activations.