Will this open model fit on my machine?

Pick an open-weight model and your hardware, and see whether it fits in memory at each quantization level, with the math shown. Parameter counts come from the official model repos.

Model Quantization Context length (tokens) KV cache precision Hardware

GPU / unified memory (GB) System RAM (GB) Unified memory (CPU and GPU share one pool)

Assumptions you can change

Usable share of memory %

The OS, display and other apps take some memory. On Macs the GPU can only use part of unified memory by default. This is your assumption, not a measurement.

Runtime overhead (GB)

Activations, buffers and the runtime itself. Varies by runtime and batch size.

KV cache override (GB, optional)

The math

Every quantization level for this model

Format	Bits / weight	Weights	KV cache	Total	On your hardware

How it works

Weights = parameters × bits per weight ÷ 8. Bits per weight come from each format's block layout (for example GGUF Q4_K stores 256 weights in 144 bytes = 4.5 bits).
KV cache = 2 (K and V) × layers × KV heads × head dim × tokens × bytes, using each model's published config. Sliding-window layers only cache the window. Compressed (MLA) caches use the latent size. Where the architecture isn't covered, we say so instead of guessing.
Overhead is a flat figure you can change.
Mixture-of-experts models need all experts in memory (or memory-mapped from disk). The active parameter count affects speed, not how much memory the weights take.
These are memory estimates only. Whether it runs at a usable speed depends on memory bandwidth, the runtime and offloading, and we don't estimate tokens per second.

Model table

Hand-maintained. Each row links to its official repo, checked . "files" = exact tensor count in the published safetensors; "card" = figure stated on the model card.

Model	Total params	Active	Source	License (per repo)

Pricing

Proposal, not final.

Free

The calculator, every model in the table, shareable links.

API (proposed $4/month)

A small sizing API (same math as this page) and hardware shopping lists for a target model. Built if there's demand.

API: coming soon