breadlab-01 · free tool estimates, not gospel — see notes below
Local AI tool

Can I run it?

Pick a model, a quantization, and your hardware. Get an honest estimate of whether it fits in memory and how many tokens per second to expect. Works for GPUs, Apple Silicon, DGX Spark, Strix Halo, and plain CPUs.

est. tokens / second

■ weights ■ KV cache ■ overhead

math runs in your browser — nothing is uploaded

How the estimate works

Fit: weights ≈ parameters × bits-per-weight ÷ 8, plus ~10% runtime overhead, plus a KV cache that grows with context length. Unified-memory machines (Macs, DGX Spark, Strix Halo) reserve a slice for the OS, so usable memory is less than the sticker number.

Speed: generation is memory-bandwidth-bound — every new token streams the whole model through memory. Estimated tok/s ≈ bandwidth ÷ model size × ~65% real-world efficiency. Prompt processing is compute-bound and much faster; this estimates generation.

Honesty: these are planning estimates for dense models. Mixture-of-experts models run faster than their size suggests. We validate this math against real runs on real hardware in Lab Notes — published methodology, raw numbers.

Built in BreadLab

This tool is free, runs in your browser, and is maintained by a real homelab in Spartanburg, SC. We also host game servers and websites.