你有没有算过,自己每天花多少时间打字?微信回复、邮件撰写、文档编辑……我们把大量精力花在"把脑子里的想法敲进屏幕"这件事上。说话明明比打字快得多,但长期以来,语音输入一直是个"鸡肋"——说一句话错几个字,改来改去还不如手打。
直到AI大模型来了,一切开始改变。
📌 快讯
千问输入法macOS版正式上线
6月27日,阿里旗下千问输入法macOS版正式上线官网,号称"说完即成稿",最快300字/分钟,支持AI自动润色和9种方言,关键是——纯净无广告。
这不是千问第一次尝试输入法。今年5月,千问曾在PC端上线AI语音输入组件,但那更像是千问App的功能延伸。而这一次,千问做了一个独立的输入法产品,直接对标搜狗、讯飞、豆包。
更值得关注的是,BAT三巨头——字节跳动、阿里巴巴、腾讯——已经全部入局AI输入法赛道。一场关于"AI时代第一入口"的争夺战,正在打响。
🔵 核心能力
四大核心卖点
① 极速输入:最快300字/分钟
普通人打字速度大约在40-80字/分钟,千问的300字/分钟意味着你正常语速说话,它就能跟上节奏。说完一段话,屏幕上已经是工整的文字。
② AI自动润色:不只是听写,而是理解
这是千问和传统语音输入最大的区别。传统语音输入是"听写员",你说什么它就转什么,口误、语气词全都保留。而千问会自动过滤"嗯""啊""那个"等口水词,纠正口误和同音字,根据上下文智能断句加标点,将口语化表达转化为书面语。
举个例子,你说:"王总,那个会议定在两点……不对三点,地点在老会议室。"传统语音输入会原样转写,而千问直接输出:"王总,会议定在三点,地点在老会议室。"
③ 支持9种方言
对不习惯说普通话的用户友好。虽然数量不及讯飞的202种,但主流方言场景已经够用。
④ 纯净无广告
输入法作为高频工具,广告一直是体验痛点。千问承诺无广告,对被传统输入法广告困扰的用户吸引力不小。
使用方式很简单:Mac用户按下右Command键唤醒语音输入,支持在微信、钉钉、邮件、浏览器等任意应用中使用。
🔬 技术解读
从"听写员"到"理解者"
过去几年,语音输入经历了两次质变。
第一次是深度学习替代传统模型,识别准确率大幅提升,但仍然"错几个字就要手动改"。第二次是大语言模型重塑语音输入——模型不再只是"听到什么转什么",而是理解你说的是什么意思,然后输出最合理的文字。
大模型解决了三个核心问题:理解语义自动纠错、智能过滤冗余表达、输出可直接使用的书面语。
千问的核心技术支撑来自Qwen大模型,采用端云结合架构:本地部署轻量级感知模型负责实时语音转写,云端由Qwen大模型负责语义理解和润色,实现低延迟的"说完即成稿"体验。
这意味着,语音输入已经跨过了"偶尔应急"的阶段。在走路、快速回复、整理想法时,说话正在变得比打字更高效。
🤔 深度洞察
为什么巨头都在重新做一遍输入法?
输入法早就是一个成熟品类,大厂为什么还要重新做?
答案是:输入法是离用户意图最近的入口。
无论你在淘宝搜索、钉钉回复、微信聊天,只要你想把想法交给手机,输入法就会出现。它不属于某一个App,却能跨过几乎所有App。
大模型让输入法从"工具"变成了"助理"——不再只是帮你打字,而是能理解你想说什么、该怎么说。
对BAT来说,战略价值各不相同:字节要把豆包能力带出独立App;腾讯要连接微信社交生态;阿里则有机会让千问穿过电商、支付、办公,从"需要主动打开的AI助手"变成"随时可调用的底层能力"。
谁掌握了输入法,谁就掌握了用户意图的第一道关口。
⚠️ 冷静思考
局限不能忽略
• 平台覆盖不足:目前仅macOS,Windows和手机用户暂时无法使用
• 方言支持有限:9种 vs 讯飞202种,差距明显
• 跨设备同步缺失:词库、常用语无法手机电脑同步
• 语音输入天然局限:办公室不方便开口,密码和精确编辑仍需手动
语音输入是打字的重要补充,但短期内还无法完全替代键盘。
💬 结语
千问输入法macOS版的上线,证明了一件事:大模型正在重新定义"输入"这件事。 输入法不再只是把声音变成文字的工具,而是能理解你、帮你表达的智能助理。
赛道才刚刚开始。BAT各有优势,也各有短板。最终谁能胜出,不取决于谁的模型最强,而取决于谁的产品体验最好、谁的生态覆盖最广。
对普通用户来说,这是最好的时代——选择越来越多,体验越来越好,而且免费。
📢 觉得有用?
回复【工具包】领取我整理的职场AI工具清单
Have you ever calculated how much time you spend typing every day? Replies on WeChat, drafting emails, editing documents — we spend a huge amount of energy on "typing our thoughts into the screen." Speaking is clearly much faster than typing, but for a long time, voice input has been "useless" — say one sentence and get several words wrong, and fixing them is slower than just typing.
Then AI large models arrived, and everything began to change.
📌 Breaking News
Qianwen Input Method for macOS officially launched
On June 27, Ali's Qianwen Input Method for macOS was officially released on the official website, claiming "instant transcription after speaking," with speeds up to 300 words/minute, AI automatic retouching, and support for 9 dialects — crucially, pure and ad-free.
This isn't Qianwen's first attempt at input methods. In May this year, Qianwen launched an AI voice input component for PC, but that was more like an extension of the Qianwen app. This time, Qianwen created an independent input method product, directly competing with Sogou, iFly, and Doubao.
What's more significant is that the BAT giants — ByteDance, Alibaba, and Tencent — have all entered the AI input method race. A battle for the "first entry point in the AI era" is unfolding.
🔵 Core Capabilities
Four key features
① Lightning-fast typing: Up to 300 words/minute
The average typing speed is 40–80 words/minute. Qianwen's 300 words/minute means it can keep up with normal speech — you say a paragraph, and the screen shows the polished text instantly.
② AI automatic retouching: Not just transcription, but understanding
This is where Qianwen differs from traditional voice input. Traditional systems are "transcribers" — they write down exactly what you say, including mistakes and filler words. Qianwen filters out "um," "ah," and "that," corrects homophones and errors, and intelligently punctuates based on context, turning spoken language into formal writing.
For example, say: “Wang zong, that meeting is at two… no, three o’clock, in the old conference room.” Traditional input would transcribe verbatim; Qianwen outputs: “Wang zong, the meeting is at three o’clock, in the old conference room.”
③ Supports 9 dialects
Helpful for those who don’t speak standard Mandarin. While fewer than iFly’s 202, the major dialects are covered.
④ Pure, no ads
Ads are a pain point in high-frequency tools like this. Qianwen promises no ads — appealing to users tired of traditional input method ads.
To use: Press the right Command key on Mac to trigger voice input, works in WeChat, DingTalk, mail, browsers, etc.
🔬 Technical Insight
From "transcriber" to "understander"
Voice input has undergone two major shifts.
First, deep learning replaced traditional models, boosting accuracy — but still requiring manual correction. Second, large language models redefined voice input: models now understand what you meant and output the most appropriate text.
Large models solve three core problems: automatic error correction, redundancy filtering, and usable written output.
Qianwen leverages the Qwen model and a hybrid architecture: local lightweight model for real-time transcription, cloud-based Qwen for semantic understanding and retouching, ensuring low-latency "instant drafting."
This means voice input is no longer just for emergencies — it’s becoming more efficient than typing for quick replies and note-taking.
🤔 Deep Analysis
Why are giants re-creating input methods?
Input methods are mature — why re-do them?
Answer: They’re the closest to user intent.
Whether searching on Taobao, replying in DingTalk, or chatting on WeChat, whenever you want to share an idea, an input method appears. It’s not tied to one app, but spans almost all.
Large models turn input methods from “tools” to “assistants” — not just typing, but understanding and helping express.
For BAT: ByteDance wants to bring Doubao out of its app; Tencent connects to WeChat; Ali aims to make Qianwen ubiquitous across e-commerce, payments, and office suites — from “active assistant” to “always-available capability.”
Who controls the input method controls the first gate to user intent.
⚠️ Sobering Notes
Limitations exist:
• Platform coverage: Only macOS — no Windows or mobile yet
• Limited dialect support: 9 vs. iFly’s 202 — still catching up
• No cross-device sync: Wordlists, favorites don’t transfer between phone and PC
• Inherent speech limits: Can’t speak in quiet offices; passwords and precise editing still need keyboard
Voice input complements typing — it won’t replace keyboards anytime soon.
💬 Conclusion
Qianwen Input Method for macOS proves: large models are redefining “input.” It’s not just sound-to-text — it’s an intelligent assistant that understands and helps you articulate.
The race is on. BAT each have strengths and gaps. Victory depends not on who has the best model, but who delivers the best product and ecosystem coverage.