How it works

Everything runs in your browser. Nothing about your machine is uploaded.

Detection

We read the WebGL renderer string, navigator.deviceMemory, navigator.hardwareConcurrency, and the user agent. VRAM comes from a GPU database, unified memory on Apple Silicon, or an estimate from RAM. Chromium caps deviceMemory at 8 GB, so the chip is cross-checked on Apple machines.

Ranking

Models are filtered to those that fit in your VRAM at Q4, then ranked by capability against the resource cost. The trade-off slider leans the ranking toward speed or toward capability. Tight fits are marked because a model that barely fits has less room for context.

Speed

Speed is an estimate, not a benchmark. Decode is bandwidth-bound, so the estimate uses your GPU or chip memory bandwidth, a backend efficiency coefficient, and the active weights of the model. Sparse models only read their active parameters per token. The result is shown as a band, and the terminal is a simulation at that speed.

Cost

Local is a fixed cost plus electricity. Cloud is a marginal price per token. The calculator shows a frontier model and DeepSeek side by side because the gap between public prices is the useful part. Cloud prices come from models.dev, with four input tokens per output token.

Data

Model data comes from models.dev and the Auxot model registry, refreshed by a weekly job. GPU bandwidth comes from GPU Ark. Apple bandwidth is curated from Apple's spec pages. The credits page lists the sources.

How detection works

GPU
WebGL renderer string
VRAM
GPU database lookup, unified memory on Apple, or estimate from RAM
RAM
navigator.deviceMemory, cross-checked against the detected chip on Apple Silicon
CPU
navigator.hardwareConcurrency
OS
user agent parsing
WebGPU
navigator.gpu presence