Find the right model, hardware & setup for your workload.

LLM Configurator is a free decision engine for local AI. It checks which of 169 open-weight model variants fit your GPU or Mac, estimates how fast they run, compares hardware and costs, and links 114 setup guides. Every VRAM figure and speed estimate comes from one shared calculation, so the same question gets the same answer on every page. Start with the GPU & VRAM checker.

It depends on memory first: a model runs well only when its weights, context cache and runtime overhead fit in your GPU's VRAM or your Mac's usable unified memory. Pick your card and we rank every model that fits, with the quantization used and an estimated speed. Models that only fit by spilling into system RAM are listed separately, because they run much slower. Browse ready answers on the Can I Run It? pages.

Roughly the model's weights plus a context cache plus a little runtime overhead. Weights scale with total parameters and quantization: Llama 3.1 8B needs about 16 GB of weights at FP16 but only 4.8 GB at Q4_K_M. Mixture-of-experts models need memory for ALL their parameters, not just the active ones. Longer context adds more. Our VRAM requirements guide works through the numbers by model size.

Memory fit is calculated from each model's published parameter count, the quantization's bits per weight and a context cache, then compared with the usable memory of your card; it is accurate to within the runtime overhead the estimate assumes. Speed is harder: it is a calibrated estimate shown as a range, not a promise. The methodology page publishes every formula and constant we use.

Mostly estimated, and always labelled. We cite 14 measured runs across 3 GPUs, each credited to the third party that published it; everything else is an estimate from a latency model calibrated against those runs, shown as a range rather than a single number. When reader-submitted runs are reviewed, they appear labelled as community results. See the benchmark database for which is which.

Choose by memory first, then bandwidth. Memory decides which models fit at all, and nothing else can compensate when a model does not fit. Among cards with enough memory, bandwidth decides how fast tokens come out. Price per gigabyte of VRAM matters more than gaming performance. Compare every card on the GPU comparison table.

Yes. Apple Silicon Macs share one pool of unified memory between the CPU and GPU, so a Mac with a lot of memory can load models no single consumer graphics card can hold. Only part of that pool is available to the GPU by default; we size models against about 75% of it. Bandwidth varies a lot between chips, and it sets the speed. Our Mac buying guide compares chips and memory sizes.

Start from the job, not the leaderboard. Coding, agents, document search, writing and translation reward different models. Then filter by what fits your memory at a sensible quantization, and prefer the strongest model that still leaves room for the context length you need. A smaller model that fits comfortably usually beats a bigger one squeezed into RAM. Our best models by workload pages do this ranking for you.

VRAM is memory soldered to a graphics card; only the GPU uses it, and it is very fast. Unified memory, on Apple Silicon and some AI mini-PCs, is one pool shared by the CPU and GPU, so the GPU can use far more of it, but the operating system and apps need a share too. For local AI, compare usable memory and bandwidth rather than the label. The hardware database lists both for every device.

Yes. The build generator starts from what you want to run and your budget, then recommends a GPU, RAM and storage combination, with the models each configuration can actually load. It also covers compact AI stations and laptops if a tower is not what you want. Each recommendation links to its compatibility results, so you can check the fit before you buy. Try the build generator.

Yes. Once a model is downloaded, runtimes such as Ollama, LM Studio and llama.cpp run it with no internet connection at all, and your prompts never leave the machine. That also works on phones with offline apps, within much tighter memory limits. For a fully air-gapped setup, download models and installers on a connected machine first, then move them across. Our offline and air-gapped guide covers the steps.

Our guides cover the runtimes most people use to run models locally: Ollama, LM Studio, llama.cpp, vLLM for high-throughput serving, and MLX on Apple Silicon. Memory and speed estimates assume GGUF quantizations as these runtimes load them; vLLM and MLX use their own formats, so treat those figures as a guide. Our Ollama vs LM Studio comparison is a good place to start.

About 66.6 GB at Q4_K_M, before context. Scout is a mixture-of-experts model with 17B active and 109B total parameters: only the active experts are read per token, which keeps it fast, but all 109B must sit in memory. That is why it needs a large unified-memory machine or a data-centre GPU, and why Maverick needs about 242.3 GB. See the Llama 4 Scout model page for every quantization.

The full DeepSeek R1 is a 671B mixture-of-experts model that needs about 405.9 GB at Q4_K_M, which means multi-GPU servers or very large unified-memory machines. The R1 Distill models are different, much smaller models trained to imitate it: the Llama 8B distill needs about 5.6 GB and the Qwen 32B distill about 20.1 GB. Compare them on the DeepSeek R1 model page.

Quantization stores model weights at lower precision, which shrinks the memory they need. Llama 3.1 8B needs about 16 GB of weights at FP16 and about 4.8 GB at Q4_K_M, with a small quality cost. Q4_K_M is the usual balance of size and quality; Q8_0 is close to lossless; below Q4 quality drops noticeably. Read quantization explained for the trade-offs.

Yes. llama.cpp, and the tools built on it such as Ollama and LM Studio, run models on the CPU with no graphics card at all. It is much slower, because system RAM has far less bandwidth than VRAM, so small models are the practical choice, and you need enough RAM to hold the whole model. The checker handles a machine with no GPU. Our beginner's guide walks through a first install.

It depends on how much you use it. Local AI costs the hardware, spread over its useful life, plus electricity for the hours it runs; cloud AI costs per token or per subscription. Heavy, steady use tends to favour local hardware, and light use tends to favour the cloud. Privacy and offline use can matter more than price. Put in your own numbers with the local vs cloud cost calculator.