Free weights for programmers.

How it works

Red → green → stronger.

1. 01 Pick a rep.

   Three levels — Foundations, Core, Principal — 8 chapters, 24 reps per language track. Every rep opens red: the tests are written, the function is empty.

2. 02 Make it green.

   A Monaco editor and real runners: TypeScript executes in a sandboxed blob worker with a 4-second kill switch; Python runs on Pyodide (CPython compiled to WASM) with 10. No simulation for these two — the test output is the truth.

3. 03 Get coached.

   Qwen2.5 Coder reviews your working diff — pick your weight class, 0.5B to 7B, downloaded once through web-llm and cached. The coach runs on your GPU; the conversation never leaves the tab.

4. 04 Come back.

   Spaced repetition resurfaces finished reps at 1, 3, 7, and 14 days; timed drills and the Rosetta library (12 patterns × 4 languages) keep the warm-up honest.

The specimen

One real rep, both states.

ts-values · Values & Types — the first Foundations rep, exactly as the gym serves it.

// the starter — red by construction
function celsiusToFahrenheit(c: number): number {
  throw new Error("TODO: implement celsiusToFahrenheit");
}

✕ celsiusToFahrenheit converts freezing point
  TODO: implement celsiusToFahrenheit
✕ celsiusToFahrenheit converts boiling point
  TODO: implement celsiusToFahrenheit
✕ celsiusToFahrenheit handles decimals
  TODO: implement celsiusToFahrenheit

// the lift — two lines
function celsiusToFahrenheit(c: number): number {
  return (c * 9) / 5 + 32;
}

✓ celsiusToFahrenheit converts freezing point
✓ celsiusToFahrenheit converts boiling point
✓ celsiusToFahrenheit handles decimals

Guarantees

Works offline

model + Pyodide cached after first download; everything else is static

No account, no API key

there is no server to hold either

Code never leaves the device

all state in localStorage under freeweights:*

Honest runners

TS/Python execute for real; Java/Rust wear an “AI estimate” badge