Inside LLM

Ever wondered what your keyboard is doing when it suggests the next word? Inside LLM runs a real language model on your iPhone or iPad and draws every step of its thinking. This is not a chatbot and not an online service. The app contains its own transformer inference engine, written from scratch in Swift on Apple's Accelerate framework, and computes the full network on your device after every keystroke so that every intermediate number can be shown. Type anything and see: - Suggestions: the three most probable next words, exactly as a predictive keyboard would show them, with their real probabilities. - Signal through the network: six layers lighting up as the forward pass runs. - Attention: a heatmap of which earlier words each token is reading from, for any of the 72 attention heads, or averaged. - Neurons: all 3,072 MLP neurons in each layer, firing for your last word. - Logit lens: what the model would guess if it stopped after layer 0, 1, 2, and how that guess sharpens or changes as it goes deeper. - Weights: every one of the model's 76 weight tensors, rendered as a zoomable heatmap of the actual numbers learned in training. - Steps: a six-step walkthrough of one forward pass (tokenize, embed, attend, MLP, stack, softmax) using your own words as the example. Everything runs locally. No account, no analytics, no ads. Nothing you type leaves the device. An optional tip jar in About lets you support development; tips unlock nothing and every feature is free. The model is DistilGPT2 (Hugging Face, Apache 2.0), an 82-million-parameter distilled GPT-2 trained on 2019 web text. Its suggestions are shown for learning purposes and can be wrong or dated. The engine's output is checked against an independent reference implementation in the app's test suite, and the method is described in an open technical report.