How does the build planner pick a GPU?

At a fixed 8K context with fp16 KV, it first searches Q4_K_M for the cheapest in-budget setup that fits the model: single card first, then 2–4 identical cards if no single card fits. If the whole tier fails it retries at Q3_K_M, then Q2_K. If all three fail, the budget is declared insufficient.

What class of model can a $1,500 budget run?

At current reference prices, $1,500 buys a used RTX 3090 (24GB, ~$1,100), which runs Llama 3.1 8B and Qwen3 8B comfortably at Q4_K_M and 27B–32B models tightly. A 70B model needs a dual-card setup starting around $3,000.

Why does it recommend a single card over multi-GPU?

At the same budget a single card is cheaper to power, simpler to run, and avoids the 0.85 multi-GPU bandwidth penalty. Identical 2–4 card setups are only recommended when no single card fits the model; usable VRAM across cards is summed without discount. Macs never participate in multi-card plans.

How accurate is the electricity estimate?

It assumes TDP × 50% average load × 720 hours × $0.15/kWh — the same convention as our cost comparator. Inference is a bursty workload, so real bills are usually lower; enter your monthly usage to see the electricity cost per million tokens.

Build Planner

Enter a budget and a target model; get the cheapest build that fits: GPU and count, quantization, VRAM breakdown, speed, electricity, and a llama.cpp command.

Budget (USD): $1,500Target modelMonthly usage (M tokens, optional)

Recommended build

RTX 3060 12GB

Quantity: 1 card(s)
Unit price: $260 used
Total: $260
Recommended quant: Q4_K_M
Theoretical speed: ~137 tok/s
Est. monthly electricity: $9.18

VRAM usage

Estimated at 8K context, fp16 KV, 1.5GB runtime overhead

Weights: 2.0 GB
KV cache: 0.9 GB
Runtime overhead: 1.5 GB
Total required: 4.4 GB
Usable VRAM: 12 GB

Theoretical estimate ±30%; real speed varies with framework, drivers and CPU · Electricity assumes TDP × 50% load × 720h × $0.15/kWh

llama.cpp launch command

llama-server -m llama-3.2-3b-Q4_K_M.gguf -ngl 99 -c 8192

Alternatives

RX 7900 XTX

Faster

Quantity: 1 card(s)
Unit price: $650 used
Total: $650
Recommended quant: Q4_K_M
Theoretical speed: ~366 tok/s
Est. monthly electricity: $19.17

Theoretical estimates: speed ±30%; prices are manual reference prices (used price when available); electricity assumes 50% average load.

How to use this planner

Drag the budget slider or type an amount ($200–$20,000). The budget covers the GPU (or whole Mac), not peripherals like monitors or PSU upgrades.
Pick a target model.The list covers all 14 models on this site, including MoE models (VRAM counted on total parameters) and DeepSeek-R1 (MLA formula).
Read the build sheet.The recommended plan shows GPU model and count, unit and total price, the recommended quantization tier, a VRAM breakdown (weights / KV / overhead vs usable), theoretical speed, estimated monthly electricity, and a ready-to-adapt llama.cpp launch command.
Optionally enter monthly usage.With a token volume filled in, the sheet adds an "electricity per M tokens" line so you can reconcile against API pricing.

The recommendation order

The planner does not brute-force every combination — it follows a fixed order looking for the cheapest in-budget setup. It starts at Q4_K_M (8K context, fp16 KV): single cards that fit comfortably or tightly, cheapest first; if no single card fits, each discrete GPU is tried as 2–4 identical cards (usable VRAM summed with no discount) and the cheapest combination within budget wins. Only when the entire tier fails does it fall back to Q3_K_M, then Q2_K. If all three tiers fail you get an "insufficient budget" verdict with links to smaller models and cloud GPU rental. A tight Q4 beats a comfortable Q3 — quality loss matters more than headroom, the same rule our compatibility checker uses.

Mac unified-memory machines are derated to 75% of nominal RAM before the fit check and never join multi-card plans; expensive datacenter cards like the A100/H100 are naturally filtered out by the budget cap. The alternatives block shows one or two counterpoints: "cheaper" is usually the same card at a lower quant or a lower-tier card, "faster" is the in-budget option with higher throughput, so you can trade price against speed.

How each number on the sheet is computed

VRAM = weights (parameters × quant bpw ÷ 8, with bpw reverse-engineered from real GGUF file sizes) + KV cache (GQA or MLA formula, fp16) + a fixed 1.5GB runtime overhead; usable VRAM is nominal for discrete cards and nominal × 0.75 for Macs. Theoretical speed = memory bandwidth × 0.75 ÷ per-token weight reads, with a further 0.85 multi-GPU penalty, labeled ±30%. Electricity = TDP × 50% average load × 720h × $0.15/kWh, the same convention as our cost comparator; prices are manually maintained reference prices (used price when available). The llama.cpp example uses -ngl 99 for full GPU offload and adds --split-mode layer for multi-card plans.

FAQ

How does the build planner pick a GPU?: At a fixed 8K context with fp16 KV, it first searches Q4_K_M for the cheapest in-budget setup that fits the model: single card first, then 2–4 identical cards if no single card fits. If the whole tier fails it retries at Q3_K_M, then Q2_K. If all three fail, the budget is declared insufficient.
What class of model can a $1,500 budget run?: At current reference prices, $1,500 buys a used RTX 3090 (24GB, ~$1,100), which runs Llama 3.1 8B and Qwen3 8B comfortably at Q4_K_M and 27B–32B models tightly. A 70B model needs a dual-card setup starting around $3,000.
Why does it recommend a single card over multi-GPU?: At the same budget a single card is cheaper to power, simpler to run, and avoids the 0.85 multi-GPU bandwidth penalty. Identical 2–4 card setups are only recommended when no single card fits the model; usable VRAM across cards is summed without discount. Macs never participate in multi-card plans.
How accurate is the electricity estimate?: It assumes TDP × 50% average load × 720 hours × $0.15/kWh — the same convention as our cost comparator. Inference is a bursty workload, so real bills are usually lower; enter your monthly usage to see the electricity cost per million tokens.