Plan inference & fine-tuning — hardware, VRAM and cost.
Compare open-source LLMs across a catalog of hundreds of models; live cloud prices, TCO and per-token cost analysis. No sign-up, free.
Start Calculating →Scenarios and numbers
These cards are computed server-side by the same calculator engine; each is editable in the app.
Inference
TTFT, TPOT, tokens/s and VRAM — from 8B to 671B MoE.
Fine-Tuning
GPU hours, VRAM and platform cost for QLoRA, LoRA and full fine-tuning.
Live Cloud Prices
Up-to-date GPU prices from RunPod, Lambda and Modal; on-prem TCO comparison.
How the calculation works
VRAM requirements
Weights = parameters × bytes-per-weight (quantization), plus KV cache, activations and runtime overhead.
Speed
Time to first token (TTFT) is compute-bound; each following token (TPOT) is memory-bandwidth-bound. Tokens/s per user and total system throughput are computed.
Cost
GPU count × hourly price × time; per-1M-token cost and price comparison across cloud providers.
On-prem TCO & fine-tuning
Hardware (CAPEX) + electricity/cooling/maintenance (OPEX) and a break-even point vs. cloud. For fine-tuning: method/framework + dataset + GPU hours + platform cost.
Data sources & transparency
Live GPU prices
GPU prices are scraped periodically from RunPod, Lambda and Modal and shown as live cloud rates in the app. There is no sponsorship, partnership or affiliation with these providers — the tool is independent and free.
Models
Models are fetched from the Hugging Face Hub (open-source, top-by-downloads + a Turkish allowlist) on top of a hand-curated catalog; architecture data is verified against HF where possible.
Hardware coverage
Includes Apple Silicon (local devices) and custom GPU/model support; local hardware has no cloud price.
Calculations are estimates; verify before making a purchase decision.
Models are fetched live from the Hugging Face Hub; GPUs cover NVIDIA/AMD/Apple.