Local AI PC Builder: Pick the Right Hardware for Your LLM Workflow

Tell us the model size, workflow, stack, budget, and operating system you want to run. The builder returns three practical build tiers — minimum viable, recommended, and no-compromise — with the GPU, VRAM, RAM, CPU, and runtime path for each. Pick “Upgrade current PC” and it asks what you already own, then tells you what to keep, what to buy, and what to verify before buying anything.

Start planning ↓

Quick-start examples

Target model classThe mainstream local chat and general-purpose tier.Primary workflowGeneral chat, drafting, and everyday assistance.Preferred stackBudget rangeHardware preferenceOperating systemBuild type

This planner uses deterministic rules, not live benchmarks or live pricing. It suggests hardware classes and qualitative speed expectations to narrow your decision. Typical price ranges, where shown, are periodically checked editorial estimates dated at the point of the figure, never live quotes or offers — always confirm current specs and prices before buying. No accounts, no email capture — builds you save stay in this browser on this device only. Result cards may include a clearly labeled retailer availability link, which never changes what we recommend.

This is a planning tool that uses practical rules, not live benchmarks or pricing. Treat the result as a shortlist to research before buying. Already own a machine? Choose “Upgrade current PC” above to plan around it — or check what it runs today:

Check what your current PC can run →Browse the hardware directory →What is VRAM and why it matters →

How the builder decides

Show the decision factors

VRAM first

Memory decides what fits

The amount of GPU VRAM (or Apple unified memory) sets the ceiling on which models you can load. Everything else tunes speed and comfort.

Bandwidth & CPU

They set the speed

Memory bandwidth and CPU affect how fast a model responds, not whether it runs. RAG and serving lean harder on RAM and storage.

CUDA vs unified

Two honest platforms

NVIDIA CUDA gives the widest tooling and best speed per gigabyte. Apple unified memory fits larger models in one pool with narrower runtime support.

No overpromises

Realistic on 70B and MoE

Large 70B and frontier MoE models need workstation memory, multi-GPU, high unified memory, or cloud — the builder says so instead of pretending an 8GB card is enough.

Who the PC Builder is for and how to read it

Show who this is for

The Local AI PC Builder answers “what should I actually buy to run this?” — either from scratch or from the machine you already have. Describe the local AI you want to run and it works backward to the hardware. Choose “Upgrade current PC” and it also asks what is in your case today, then sorts every part into keep, buy, or verify: anything you already own costs nothing, and anything the tool cannot know — your power supply’s wattage, your case’s clearance — is returned as a check to make rather than a number we invented. Read the three tiers as a spectrum, not a ranking: “minimum viable” is the cheapest configuration that will genuinely run your target, “recommended” is the comfortable everyday choice, and “no-compromise” leaves headroom for larger models and longer context. Treat every tier as a research shortlist to price out and verify — it plans the build, it does not place the order.