VRAMFit: LLM Calculator

Descrizione

Before you download gigabytes of model weights, know whether they'll actually fit. VRAMFit tells you the largest local large language model you can run on your hardware — whether that's an NVIDIA GPU's VRAM or the unified memory on an Apple Silicon Mac. Set your memory, pick a quantization, and get a clear, instant answer. Built for people who run models locally: llama.cpp, Ollama, LM Studio, and anyone deciding what GPU to buy next. MEASURED, NOT GUESSED Most calculators multiply a parameter count by a rule of thumb. VRAMFit ships the real size of the published file for every model at every quantization it is actually released in, and names the exact build each figure came from — so you can check it against the download before you start it. That matters most where rules of thumb break. A ternary model has no 4-bit build to estimate. A natively 4-bit model barely changes size as you move up the scale. VRAMFit says what those models actually weigh, or tells you plainly that the build you picked does not exist. WHAT IT DOES • Enter your VRAM (or Apple Silicon RAM) and instantly see what fits • Sorts the catalog into what runs, what runs one setting down, and what won't • Accounts for quantization — from full precision down to 4-bit and FP4 • Factors in context length and KV cache, the memory costs people usually forget • Switches between discrete GPU and Apple Silicon unified-memory math FREE • The full calculator, with quantization, context, and FP4 • A starter set of models spanning what fits and what nearly fits • GPU and Apple Silicon modes VRAMFit Pro (one-time upgrade) • The complete catalog of 73 current open models, measured • Per-quantization memory breakdown for every model • Advanced tools: KV-cache compression and multi-token prediction estimates PRIVATE BY DESIGN VRAMFit runs entirely on your device. No account, no sign-in, no tracking, and nothing about you leaves your iPhone. The numbers you enter stay with you. Stop guessing whether a model will fit. Check first, download once.

Model sizes are now measured, not calculated. VRAMFit used to size a model by multiplying its parameters by a rule of thumb for each quantization. It now ships the real size of the published file for every model at every setting it is actually released in, and names the exact build each number came from. • A much larger catalog of current open models, every one measured • Models that only exist at one quantization say so, instead of showing a size for a build that was never published • Natively 4-bit models no longer appear to grow as you move up the scale • The list is grouped into what fits, what fits one setting down, and what won't — so a model that just misses tells you which setting would run it • Rewritten help explaining where the numbers come from and what they still cannot know Estimates are higher than before at every setting. That is the correction: the old rule of thumb ran low.

Informazioni

Venditore

Sarvesh Anand Kandala

Categoria

Developer Tools

Versione

1.1.0

Richiede

iOS 18.0+

Dimensione

1.4 MB

Classificazione per età

4+