Find local AI models that should realistically fit your PC.

Enter your CPU, GPU, Graphics Card Memory (VRAM), System RAM, and target workflow. OpenSourcesAI will suggest good, better, and best local model profiles to try, plus a practical app/runtime path for testing them on your machine. It doubles as a VRAM calculator: give it your graphics card memory and it shows which model sizes and quantization levels fit.

Start checking ↓

Optional helper

Auto-detect my device basics

We ask your browser for approximate memory, CPU thread count, and graphics renderer info. This is not a full PC scan — review and correct any field before running the checker.

Local LLM Hardware Compatibility Matrix

Select your graphics card memory (VRAM) and system RAM. The grading engine matches your hardware against each model’s minimum requirements. We require at least 2 GB of VRAM headroom above the model’s minimum before rating a fit as Comfortable.

Your hardware

Quick-select a hardware presetPresets are optional shortcuts covering CPU-only machines, NVIDIA gaming cards, Apple Silicon, and multi-GPU workstations (the 2× and 4× GPU rigs at the end of the list). Selecting one fills VRAM, RAM, and GPU name instantly — adjust any field to fine-tune. Card not listed? Skip the presets and set VRAM and RAM directly below.

CPU — optionalOptional — helps the wizard generate better runtime notes. Does not change the match.GPU — optionalOptional context — the match is graded from VRAM and RAM. The GPU name refines Apple Silicon unified-memory handling and deployment notes.GPU VRAM in GB — requiredSelect 0 for integrated GPUs or CPU-only builds. For Apple Silicon, use your unified memory figure. Multiple GPUs? Select the combined total across cards — 2× RTX 3090 is 48 GB. Exact size not listed? Round down.

What is VRAM?

VRAM is your graphics card's dedicated memory. Larger models need more VRAM — a 70B parameter model needs roughly 42–47 GB, while an 8B model typically needs 6–8 GB. Quantization (like Q4_K_M) shrinks the model, letting smaller cards run bigger models. Learn more about VRAM →

Two GPUs, a laptop GPU, or an unusual VRAM size? How to enter it

Multiple graphics cards: select the combined VRAM total — 2× RTX 3090 is 48 GB, 4× RTX 3090 is 96 GB. Results grade against the combined pool, because runtimes like Ollama split a model across cards automatically; a note about multi-GPU speed appears with your results. The preset list also has ready-made 2× and 4× GPU entries under “Multi-GPU Workstations and Clusters.”
Using auto-detect with two cards: browsers can only see one graphics adapter, so auto-detect reports a single card — raise the VRAM figure to your combined total yourself before running the check.
Apple Silicon (M1–M4): enter your total unified memory as the VRAM figure. The checker automatically sizes against about 75% of it, since macOS shares that memory with everything else you have open.
Laptop GPU: use the card's dedicated VRAM — a laptop RTX 4060 is 8 GB. Ignore “shared graphics memory” in Windows settings; that is system RAM, which the checker already counts separately.
VRAM size not in the list: pick the next size down — an 11 GB GTX 1080 Ti checks as 10 GB. Underestimating keeps the verdict honest.
External GPU (eGPU): enter its VRAM normally. The enclosure slows model loading, not whether a model fits.

Planning a dual-GPU build? Read the dual RTX 3090 guide →

System RAM in GB — requiredIf a model is too big for your graphics card, your computer can run part of it in ordinary system memory instead — slower, but it still works. More RAM keeps more models within reach.Primary workflowWhat you want the model to do — the biggest factor in how results are ranked.

Which workflow should I pick?

Pick the task you will do most often — it carries more weight than any other answer here. Choosing coding moves models tuned for code to the top; choosing document Q&A favours models that handle long documents and retrieval well. Nothing is locked in: change it and run the check again to compare. Compare models by use case →

PreferenceNudges the ranking toward faster or higher-quality versions of each model.

Which preference should I choose?

The same model comes in different sizes on disk. Speed first favours the more compressed versions — they load quicker and leave VRAM to spare, with a small drop in answer quality. Quality first favours the fuller versions, which answer better but need more memory. Balanced sits between the two, and is the right default if you are unsure. How model compression works →

Scores use practical VRAM fit rules — not live benchmarks. Confirm specs against your GPU manufacturer before long inference sessions.

Results are ranked for: Chat and general assistant · Balanced quality and speed — change the two fields above to match your goal.

Set your GPU VRAM or system RAM above to run the check — at least one of the two is needed. Not sure? Use auto-detect or pick a preset.

This checker uses practical rules, not live performance tests. Treat the result as a starting point before testing a compressed model locally — lower-memory PCs may show only one realistic local match. See exactly how results are computed →

Next steps

No PC yet? Plan a build →Open the Model Builder Wizard →Compare local LLM tools →Browse model profiles →New to the terms? Read the quickstart →

Quick definitions

Show definitions

VRAM

Graphics-card memory

VRAM is the dedicated memory on your GPU. Common values are 4 GB, 8 GB, 10 GB, 12 GB, 16 GB, or 24 GB.

Quantized model

A compressed local model

A 4-bit quantized model uses less memory, usually with a small quality tradeoff, so it is easier to run on a normal PC.

Runtime

The app that runs the model

Ollama and LM Studio are beginner-friendly apps for testing local models. llama.cpp is a technical engine used by many local model tools.

RAG

Answers from your files

RAG means the model checks your documents or notes first, then answers using those sources.

Open-weight

Public model weights

Open-weight means the model weights are public, but the usage terms may still have rules.

Using this page as a VRAM calculator

If your question is “how much VRAM do I need to run a local LLM?”, this tool answers it from the direction that matters: enter the VRAM you have, and it calculates which model sizes and quantization levels fit — with headroom for context — instead of leaving you to work the math out per model. The fit estimates use realistic quantized model sizes rather than best-case numbers.

Want to run the numbers by hand, or estimate a specific model at a specific context length? The VRAM guide walks through the weight and KV-cache formulas this kind of VRAM calculator is built on.

How to read your checker results

Show how to read results

This tool is for anyone who wants to run AI models on their own PC but is not sure what their hardware can handle. Enter your GPU, VRAM, and RAM, and it returns model profiles sorted into good, better, and best matches for your machine, plus a runtime path to try. Read “good” as the safe, responsive pick you can rely on every day; “better” and “best” push toward larger models that still fit but may run slower or leave less room for long conversations. If you only see one match, that is normal on lower-memory PCs — it means the checker is steering you toward the size that will actually feel usable rather than one that technically loads and then crawls.