AI Setup Planner

What model and hardware are required to run X?

Answer 3 questions. Get a practical setup recommendation: model, hardware, memory fit, and a rough cost reference.

Loading planner...

How this works

1️⃣

What do you want to do?

Code, chat, image, video, vision, RAG, or voice.

2️⃣

Which model fits your hardware?

See only models that can do that task, ranked by capability and minimum RAM/VRAM.

3️⃣

Check the hardware class

See what kind of local hardware can run it, with a rough setup estimate as extra context.

What you get

🧭

Task-to-setup guidance

Pick a task and see which model and hardware class make sense.

⚡

Performance estimate

tokens/sec for your specific hardware + model combo.

🔗

Shareable plan

URL encodes your selection. Send to friends, save for later.

💾

Browser saved plans

localStorage keeps your plans across sessions on this device.

How the rough estimate works

rough estimate = hardware + build + RAM upgrade + power

power_5y = (watts × load) / 1000 × 24 × 365 × 5 × $0.10

This is intentionally approximate. Load is 70-85% for LLM inference, not idle. Electricity uses $0.10/kWh. GPU builds include a base system estimate; extra System RAM is treated as an upgrade above the base 32GB build. Mac and unified-memory devices use their SKU price instead of separate RAM pricing.

FAQ

How accurate is the cost estimate?

It is a rough planning estimate, not a live quote. Hardware prices are MSRP or typical retail, and real prices vary by region, used/new market, taxes, parts, and electricity. The main purpose is to understand model and hardware fit; price is supporting context.

Why a "full build" line on top of the hardware price?

For GPU builds, the GPU card price only covers the graphics card itself. You also need a CPU, motherboard, 32 GB of system RAM, an 850 W PSU (RTX 4090 draws 450 W under load), a case, a 2 TB NVMe SSD for model weights, and a CPU cooler. For Mac, MacBook, and AI PC systems, this is already baked in (one-box), so the line is $0.

What costs are NOT included?

For transparency, we left out: UPS (~$100-250 for 5y runtime protection), your time on setup and updates (30-80 hours over 5 years depending on hardware), a 7-10% failure reserve for parts that wear out (HBM, fans, SSD), and mid-life replacement for Pi/Jetson-class hardware with 4-year expected lifespan. For a more complete picture, see our blog post on the v2 model.

Are the performance numbers (tokens/sec) real?

Yes, from public benchmarks on Hugging Face, Ollama benchmarks, and community posts. Actual performance varies by quantization, batch size, context length, and software (llama.cpp vs Ollama vs LM Studio).

Why doesn't my hardware show up?

We only show hardware that can actually run the selected model. For example, an 8GB Mac won't show up for a 70B model. Try selecting a smaller model, or add your hardware (V1).

Where are my saved plans stored?

In your browser's localStorage. They never leave your device. To sync across devices, sign in (V1, coming).

Can I add my own hardware?

Not yet. In V1 we'll add user-submitted hardware with community voting. For now, email us with details.