homebench is a single‑command TUI that discovers the models installed on your local runner—Ollama, LM Studio, llama.cpp, vLLM, or any OpenAI‑compatible server—and runs a curated quality suite while measuring tokens/sec, time‑to‑first‑token, and memory footprint on your actual machine. No config, no API keys, no cloud.

The tool reports three core metrics: tok/s (output tokens divided by generation time), TTFT requiring only the first streamed token, and memory (resident size plus a best‑effort peak‑RSS sample). Quality is scored on 31 deterministic tasks across math, reasoning, factual recall, instruction‑following, extraction, code understanding, and optional open‑ended tasks scored by an LLM‑as‑judge.

Installation is as simple as `pip install homebench` or `pipx install homebench`. Run `homebench` for a quick fast subset, `homebench --all` to benchmark every discovered model, or `homebench --full` for the full 31‑task suite. The CLI provides flags for provider selection, custom task packs, and output formats (TUI, Markdown, JSON). Cached responses enable near‑instant re‑runs.

Custom task packs can be created in JSON or YAML, and homebench automatically saves each run in `~/.homebench/runs`, allowing diffing and history inspection.