v1.8.3 — Your conversations now save. Python actually runs. And you can watch the model think. What's new • Conversations are saved. TokenMax was keeping your chats, folders, personas and course progress in memory only, so everything disappeared when you closed the app. They now persist properly between launches, and if storage is ever unavailable the app tells you instead of quietly losing your work. • A real Python sandbox. The Terminal's Python mode now takes multi-line programs: end a line with a colon, type the body, and send a blank line to run the block. Errors come back as Python's own traceback, with line numbers matching the lines you typed. • Run code from chat. Python code blocks in a reply have a Run button. It executes in the same on-device sandbox — no network, no subprocesses — and shows the output beneath the code. • Long conversations keep their thread. Past about three quarters of the context window, older turns are folded into a summary the model still sees, instead of being dropped without warning. Your transcript keeps every message, with a marker showing where the boundary falls; tap it to read the summary. • Under the hood. A new view showing what the model is really doing: its parameter count, layers and vocabulary read from the model file, how your sentence breaks into tokens, and — live while it writes, on by default — the model's own probability for every word it considers. The diagram of the forward pass is labelled as a schematic, because per-layer values cannot be read out on a phone without costing more than the answer itself. • Every claim is labelled. The walkthrough marks each panel as Measured, Derived or Schematic, so you always know whether a number was read from the model, calculated from it, or drawn to illustrate an idea. • How your phone is coping. A live look at thermal state, the app's own CPU use, memory headroom, whether layers are running on the GPU, and how fast words are actually being produced. Deliberately missing: a temperature in degrees and a GPU-usage percentage — iOS gives no app either number, and we would rather show you less than make something up. • Chat titles. Conversations are given a short name after the first exchange. This never worked before. • Free books. A shelf of textbooks whose authors put the full text online — deep learning, reinforcement learning, agents. Each opens in your browser. • Reasoning models are supported properly. If you load a model that thinks out loud, that reasoning streams into its own panel and folds away when the answer starts — and it no longer leaks into copied text, read-aloud or titles. Smaller fixes • The Terminal's input is a proper tap target, keeps your capitalisation, and no longer collides with the keyboard. • The message box returns to a single line when you are not typing in it. • Pasted multi-line code runs as written instead of being flattened into one line. • Long conversations are measured once rather than repeatedly while being fitted to the context window. • Sharing into TokenMax tells you when it doesn't work. A shared PDF or image that couldn't be read used to fail silently — the share sheet closed as though it had worked and nothing arrived. • Switching models tells you when one won't load, instead of appearing to switch and then not answering. • VoiceOver names the buttons that are just icons, including the Terminal's send button and the persona icon picker. • The live decoding panel moves at a pace you can read, instead of flickering or racing to the end. • The words-per-second figure shows current speed, not an average including the wait for the first word. • Dictation is on-device only: where that isn't possible, TokenMax says so instead of sending audio to Apple. Privacy stays the same: every chat, every command, every line of code stays on your iPhone. No accounts, no servers, no telemetry. Sealed mode in Settings still locks all network egress with one tap.