On-device models
OIOXO ships with on-device coding models. The first time you build, it looks at your hardware and picks the strongest tier your device can actually run — no choices to make, no cloud fallback behind your back.
The tiers
| Tier | Download | Runs best on | Notes |
|---|---|---|---|
| Basic | ~0.3 GB | Any device | General model, light coding help. Falls back to CPU if there’s no GPU (slower). |
| Light | ~1.1 GB | Capable GPU | A dedicated coder — the sweet spot for most laptops. |
| Pro | ~4.5 GB | Strong GPU | The strongest on-device brain. Strong GPU recommended. |
How OIOXO picks
Device detection checks WebGPU support, GPU memory, system RAM, and CPU cores, then classifies your machine (CPU only → Entry GPU → Capable GPU → Strong GPU) and recommends the matching tier. You’ll see the result as a chip — e.g. “Your device: Strong GPU · 12 GB RAM” — and a Recommended for your device badge on the matching model.
Downloads & caching
The model downloads once, with live progress, as the first step of your first build — then it’s cached by your browser and loads from disk on every later visit. No re-downloads, no install.
On-device tiers punch above their size because of the verified loop: the compiler and the live preview catch and correct mistakes. For frontier-model quality on large apps, plug in your own API key — same loop, bigger brain.
What to expect
- Generation speed scales with your GPU. On CPU-only devices, Basic works but is slow — BYOK is usually the better experience there.
- The on-device tiers excel at focused tasks: scaffolds, components, games, scripts, repairs, refactors. Very large multi-file apps are frontier-key territory.
- Once the model is cached, code generation itself needs no network at all.