About FitLLM

One question: will this local LLM fit your GPU's VRAM or your Mac's unified memory?

Computed with the open FitLLM engine — accurate per-layer KV-cache modeling, not a naive estimate. Updated 2026-10-29.

What it is

FitLLM is a free calculator for running open-weight LLMs locally. Pick a model, a weight quant, a KV-cache quant and a context length, and it shows whether the model fits an NVIDIA or AMD GPU's VRAM or an Apple Silicon Mac's unified memory, how much memory is used, and the longest context that still fits. No ads, no login, no affiliate links — the output is never for sale.

How it computes fit

Weights — from the parameter count and the bits per weight of the chosen quant (GGUF quants for llama.cpp on GPU, MLX bits on Mac).
KV cache — per layer, from the model's official config.json: real KV heads and head_dim, sliding-window layers capped at their window, hybrid/linear layers with no growing cache, MLA and compressed caches as they are actually stored.
Runtime overhead and reserve — on a GPU, 2 GB for Windows + display by default, 1.2 GB for Linux + display and 0.6 GB for Linux headless (2nd card); on a Mac, a macOS reserve of 8.0–9.0 GB depending on memory size.
Verdict — fits when the total stays within memory; tight when less than the headroom (5% on GPU, 20% on Mac) is left. Recommended max = the longest context that keeps the headroom; hard limit = the longest that fits at all.

FitLLM estimates memory fit, not speed. These are estimates — real usage varies with runtime, driver and display. The full reasoning, with numbers you can reproduce by hand: why naive VRAM calculators are wrong.

Open source and data

Engine: fitllm-engine v2.22.0 (MIT, one readable file, revision 5aa318a).
Open dataset of precomputed verdicts: /data · per-model architecture: LLM memory & architecture reference · every model × hardware page: fit pages.
Found a wrong number? Open an issue — measured runs with logs are the most useful reports.

▶ Open the FitLLM calculator

About FitLLM: a free, open-source calculator that checks whether a local LLM fits your GPU or Mac — no ads, no login, no affiliate links (about). All numbers are computed by the open-source fitllm-engine v2.22.0 (MIT, revision 5aa318a) from official model config.json values — reproduce or audit them yourself; how the math works. Estimates; real usage varies with runtime (llama.cpp / MLX / Ollama), driver and display. Found a mismatch? Report it. · FitLLM home