Alibaba's Qwen Input Method Lands on iOS: Voice-to-Polished-Draft, No Ads, No Signup
Alibaba's Qwen Input Method arrived on the iOS App Store on July 16. Its core is LLM-driven "speak and it's drafted": built on the CosyVoice speech model, up to 300 chars/min, 9 dialects, auto-removing filler words, fixing slips, reordering, and formatting.
AI input that starts from speaking
Per IThome, following its macOS version, Alibaba's "Qwen Input Method" launched on the iOS App Store on July 16 (iPhone only). It positions itself as an AI input tool that starts from speaking — "speak and it's drafted."
Voice: CosyVoice + dialects + long recording
- Voice input is built on Alibaba's in-house CosyVoice speech model, claimed at up to 300 characters per minute;
- Supports 9 dialects including Cantonese, Sichuanese, and Shanghainese, plus mixed Chinese-English recognition;
- Single recordings up to 3 minutes — for long emails, weekly reports, meeting notes.
The real differentiator: LLM polishing
Unlike traditional speech-to-text, the core is LLM-driven semantic understanding and content polishing. After you dictate, the system automatically removes filler words, fixes slips of the tongue, reorders phrasing, and paragraphs/formats the text.
Official example: a spoken line with a correction — "Boss Wang, the meeting's at two… no, three" — is fixed in real time into structured text: "Boss Wang, the meeting is at 3pm in the old conference room; please bring the market-research report." It can also break casual speech into checklists, minutes, or notes.
Keyboard and experience
The keyboard side offers 26-/9-key switching, smart suggestions, fuzzy pinyin, input correction, and a one-handed mode; tap the mic for a short press to talk, or long-press space to talk, swipe left to undo, right to send. The experience is pitched as "pure" — no ads, no pop-ups, no news feed, and no signup required for all core features.
One-line take
A pragmatic sample of "LLMs descending into a high-frequency entry point": embedding an LLM's clean-up ability into the input method — the most frequent daily surface — trading the deterministic value of "speech → structured draft" for retention, rather than piling on flashy features. For AI-app teams, it demonstrates the path of "pick the right high-frequency surface and deliver one deterministic small win."