Memory fits
YOUR OWN HARDWARE
NVIDIA GeForce RTX 4090
24 GB VRAM
THE LOCAL LLM HARDWARE MATCHMAKER
Your next model.
The right machine.
Calculate LLM VRAM and RAM. Find hardware that fits.
Make your next local AI setup an informed one.
BUILT FOROllamaLM StudiovLLM
01 Configure your model
Connected catalog
A tiny instruction model for experimenting on memory-limited devices.
4.85 effective bits per weight · 0.30 GB estimated weights
1K32K max
More room for conversation needs more memory.
One conversation, full GPU offload, FP16 KV cache.
YOUR MEMORY ESTIMATELive
1.5GiB
Recommended GPU memory
Room for your model. Space to think.
Model weights0.28 GiB
KV cache0.09 GiB
Runtime + headroom1.15 GiB
System RAM targetConservative loading allowance
16 GiBPlanning estimate, not a measured requirement.
02 Find your fit
Memory-compatible options Memory fits
UNIFIED MEMORY
Mac Studio · 64 GB
64 GB unified memory
Memory fits
ON-DEMAND COMPUTE
RunPod · A100 80 GB
80 GB VRAM
Provider links may be affiliate links. We may earn a commission at no extra cost to you. Memory fit does not guarantee runtime compatibility or performance.
What goes into the estimate?
Weight-file GB are converted to GiB (2³⁰ bytes). FP16 KV memory is 2 × layers × KV heads × head dimension × context tokens × 2 bytes. Runtime allowance is the larger of 1 GiB or 10% of weights. We then leave 10% of GPU capacity free.
How much system or Mac memory do I need?
System RAM is weights plus 8 GiB, rounded up to 8 GiB. Unified memory is the greater of GPU target plus 8 GiB or GPU target divided by 0.75, rounded up to 8 GiB. These are conservative loading and OS allowances, not measured peaks. Mac cards use this separate unified-memory target.
What should I check before buying?
Actual memory use depends on your runtime, model file, drivers, batching, and loading strategy. Catalog sizes use approximate parameters and effective bits per weight, except gpt-oss which uses rounded published MXFP4 downloads. Gemma 3 and gpt-oss cache estimates allow full context per layer, without sliding-window savings. Compare the actual file and usable memory before buying. Hardware GB are conservatively treated as decimal bytes. CPU offload, image processing and multi-GPU performance are not modeled. vLLM options support FP16 and native gpt-oss MXFP4 and exclude Macs.
What about ChatGPT, Claude, Gemini and Codex?
Hosted AI services do the model processing remotely. Gemma and gpt-oss have downloadable weights for local use. Codex is a coding tool, and its memory needs depend on the model behind it. Compare local and hosted AI.