INTERACTIVE WORKBENCH

Find the right KV-cache setup for your Mac

Choose your Mac, model, and goal. We’ll recommend the best KV-cache method, estimate memory use, and show how much conversation fits.

How this works

As an LLM generates tokens, it stores key-value pairs in unified RAM so it can remember earlier context. VeloxQuant-MLX compresses this KV cache with Apple Silicon Metal kernels, helping the same Mac handle longer conversations or larger models.

01 Your Mac

Apple Silicon Chip

Unified RAM

02 Your Model

Custom Architecture

Manual override for unlisted models.

Model size

Layers

KV heads

Head dimension

Choose Quantization Algorithm

Memory Budget for KV Cache

Benchmark Suite

Where these numbers come from (Sizing formulas & benchmarks)

Sizing formulas and recommendations run the exact same logic as the veloxquant recommend CLI tool. Benchmark curves in Step 3 are real, measured Metal kernel results. To run native profiling on your own machine, install VeloxQuant-MLX and run veloxquant benchmark.