iPhone & iPad · Free
LocalAI Academy
Learn how AI models actually run on your own hardware — and work out, before you spend anything, whether the machine you are looking at can run the model you want.
Learn it properly
9 paths and 47 lessons, each explained in plain English first, then technically, then why it matters when you are choosing hardware. 119 questions to check yourself.
Will it run?
Pick a model, a quantization and a context length, and see which machines hold it and how fast they decode. The KV cache is sized from each model's real attention geometry — so you can watch a long context halve your speed.
Compare the iron
17 desk-side machines scored by one published formula — memory fit, decode, prefill, tokens per watt, cost per token per second.
Look it up
53 terms from local inference, each with a plain-English explanation and a technical one. Quantization, KV cache, MoE, unified memory, MLA.
What it does not do
- No account. Nothing to sign up for. You are never asked who you are.
- No tracking. No analytics, no advertising identifier, no third-party SDKs.
- Nothing leaves your device. Progress is stored locally and not synced.
- No ads, and no buy links in the app.
The same numbers as the website
The app and this site read from one shared content layer and use the same inference model, so they cannot disagree. Every throughput figure is derived from memory bandwidth and compute by a formula that is published in full, including the constants that are least certain. The app currently covers 6model architectures, each taken from its own published configuration.
Figures are modelled estimates, not measurements, and the methodology page says exactly where the model is weak.