faster and cheaper. • up to 11× faster first token: reasoning now defaults to low (near.ai GLM 5.3 Flash 14.6 s → 1.3 s, Fireworks 5×). for hard math, dates or code, long-press the reply and tap think harder. • caching fixed: Fireworks GLM 5.3 Flash reused ~0% of earlier input, now up to 90% per message (it varies). • near.ai: reply signatures are checked against the attested model's key. NVIDIA GPU tokens and TLS quotes are now verified too. • deleted chats stay deleted.

on-device answers pick up where they left off: one cut short when you leave the app is asked again when you come back. • developer mode shows on-device speed: how fast the model reads your message and how fast it writes. • "turn off & retry" now retries without web search. • sending scrolls to your message in one smooth motion. • a server address that only looks like one on your own network is no longer treated as your machine. • a near.ai model relayed to another company no longer shows as end-to-end encrypted while the list loads. • fireworks now uses Kimi K3: new setups start on it, and a saved Kimi K2.6 setup moves to it and shows you the new price.

new providers: OpenRouter (hundreds of models, one key) and NVIDIA (free hosted tier). paste a key, browse, pick. • model lists come from each provider, live, with prices. one browser across all of them. • Gemma on the phone streams about twice as fast and loads in the background, so the first message doesn't wait. • near.ai's enclave hosts come from near.ai's live directory. a model with no published host asks before sending instead of sending in the clear. • first run walks through setup and lands you in the composer. • fixes: retired near.ai models heal to GLM-5.3 Flash, the web-search offer no longer leaves gaps, rejected keys say so.

privacy • your near.ai key is no longer written to the app's URL cache on disk. Every signed-in request now runs with the cache off, and the cache left by earlier versions is emptied the first time this version launches. • removing or replacing a key clears the cache as well. Found by the runtime pass of our own source audit (github.com/teemoonai/audits, client/teemoon-ios). The key was on disk inside the app's sandbox, readable only with file access to the device; it was never sent anywhere it shouldn't go.

downloads • models download in the background and survive the app being closed; a download resumes where it stopped • a download paused off wi-fi waits for wi-fi, and says so; mobile data is only used when you choose it • a download whose link expired while you were away starts over on its own chat • the transcript no longer shifts or opens a blank band as a reply lands • on-device replies stop cleanly, and a reply cut off by leaving the app resumes when you return setup • adding a cloud key is two steps: pick the provider, then its key • a setup without its key can't be sent to; the row says what it needs • the picker names the recommended model proof • one failed verification leg no longer discards another leg's success

now fully open source — the whole client: github.com/teemoonai/teemoon-ios • long chats stream and scroll smoothly • everyday proof shows what each audit found, not just that one exists • web search sends the model facts, not raw page markup • alert buttons finally match teemoon orange