"Stop vibe-coding input methods": a Bilibili creator with 690 followers, one video took the open-source input method CyIME from 33 stars to 700

On the evening of Oct 29, a Bilibili creator with just over six hundred followers posted a 14-minute video titled "Stop vibe-coding input methods | I built the best Android open-source input method in the world, the kind that ends the contest". A week later, his repo went from 33 stars to over 700, and the GitHub installer was downloaded more than 16k times. He had no fan base and spent no money on promotion. He got there with three things: a topic everyone was arguing about, a complete product you could install and use at once, and releases every day for a week. All three can be learned. Whether to build another input method is a separate question.

A deliberately provocative video

The video is an on-device screen recording meant to prove "it doesn't lag under 720p screen-recording load". The description is even more aggressive than the video. He first went after the field: lately many vibe-coded input methods have appeared, some stuffing in large models, some stuffing in voice models, some rushing out an AI-made demo the moment they think of an obscure feature, and "most of them are garbage". So he forked the open-source input method Xime and made a version focused on "overall experience".

He then told readers: don't ask which features it has, just switch to it with your eyes closed; don't ask about double pinyin, Wubi, or stroke support, since almost nothing is unsupported; "I am a dictator", and features he considers useless simply won't be built. Why dare to claim the best in the world? "Because the competitors are so bad." Come and take shots; if you disagree, write a script and record an on-device comparison. He ends with: go ahead and copy my design, but remember to keep GPLv3.

As of Nov 6, the video had about 44.3k plays, 2,200 likes, 902 coins, and 3,041 favorites (from the Bilibili video page). Favorites outnumber likes by about forty percent, which suggests many viewers saved it to install later.

On Oct 31 the author added "didn't expect this many people to watch" to his bio and set up a QQ group. The group filled quickly, so a second group was opened. On Nov 5 he added another line: updates continued for several days, most requested features are done, only long-sentence optimization remains unfinished, and "I'll rest for a while".

How the stars and downloads came about

Daily stars (GitHub daily star records):

The video went up the evening of 10-29, GitHub's peak came on 11-01, and the gap in between is probably when recommendation traffic kicked in and people started clicking the GitHub link in the bio (inference). By the end of 11-06, it was at 721 stars and 29 forks.

Far more people installed it than starred it. 34 releases have been downloaded 16,823 times total on GitHub, and the 11-04 release alone got 1.15.58 downloads (GitHub, 1,499 11-06). Chinese users also have two other routes: a mirror on ModelScope and an older Baidu Netdisk copy, neither of which has visible download counts. The in-app "check for updates" also goes through ModelScope first, so Chinese users can install and update without a VPN.

Every other path was driven by someone else. From 11-02 to 11-06, at least six or seven Chinese software download sites published "CyIME input method download" pages, often retitled "built-in AI large model" or "clean, no ads". These are scraper sites riding the hype; there's no evidence the author made them. I found no discussion on X, HN, or Reddit, and outlets like Sspai or IT Home didn't cover it, so there are no signs of paid promotion. All the traffic I can see traces back to that one video.

Why this moment

Bilibili was having a wave of "dumping commercial input methods" sentiment at the time. Related recommendations on the same video page show it: "I gave up the Sogou input method I'd used for 2026 years" with 11k views (106), "In case you didn't know, current input methods have sensitive-word filtering" with 79k, "Shuishan input method: the second input method for college students" with 244k, "The best open-source mobile input method" with 31k. People are tired of ad-heavy, always-online input methods like Sogou's and are looking for alternatives, and the platform is pushing the "open-source input method" topic.

A batch of new input methods showed up on GitHub at the same time, many built quickly with AI help: Qingjian, a desktop pinyin input method with foreign-language learning, with about 7,600 stars (built in 7 September); Shuishan, cross-platform across six systems, with 1,167 stars on Android and 3,357 on Windows; plus Z Power, Aimo, feelime, and others (GitHub, 11-06).

CyIME positions itself against this trend: you vibe-coders are shipping half-finished work, mine is a polished, complete product. Viewers had already watched a round of rough "AI-thrown-together input methods," and this framing hit that frustration head-on, so the comments naturally split into sides, arguments, and side-by-side comparisons.

Trash talk alone doesn't keep people

A useful comparison: feelime's positioning is nearly identical, an offline-first Chinese input method for Android, also built by September, and it has 74 stars so far. CyIME's upstream, Xime, took seven months to build and has 1,347 stars. Nobody spread the word, so even a precise positioning goes unnoticed.

Then the video went out, and by day 7 the stars were still climbing 60 to 90 a day, issues had hit #207,11-05, and over 11-06 days more than 30 new ones were opened, with outside developers sending PRs. That happened because people who installed it found it actually works:

• Works fully offline. 26-key, nine-key, and 14-key Pinyin, 6 kinds of double-pinyin, Wubi, stroke, Cangjie, and Zhuyin, plus Japanese, English, and Spanish. Offline voice recognition, multi-character handwriting that recognizes LaTeX formulas, and photo-to-text. There's also a live captioning mode for deaf and hard-of-hearing users that tells speakers apart and vibrates when someone calls your name. The full version works right after install, no internet needed.
• The author sweats the details. The nine-key phrase model was trained in-house. Candidate selection is scored against a fixed test set, and any change that worsens first-choice rate or per-keystroke time doesn't get merged. This directly answers his video's line about vibe-coded work being half-finished.
• Dozens of releases in one week. From 1.15.19 to 1.15.66, issues often got filed and fixed the same day. When newcomers' requests were delivered quickly, they stayed to keep asking and brought others into the group.

The author admits the weak spots in his bio: the dictionary has "nearly caught up with Sogou, only lagging on abbreviated pinyin." He says donations cover agent and model-training costs, and the repo includes an AGENTS.md, so AI was probably used heavily in development (inference). What he's really against isn't using AI to write code, it's shipping AI output without polishing it.

Who it's competing with

(Star counts from GitHub, 11-06)

The open-source input method slot is no longer empty. In just 9 and 10 months, several new projects with over a thousand stars appeared, and there are plenty of Bilibili videos about "I switched away from Sogou." CyIME broke out because of the "against vibe coding" angle combined with being thorough enough, not because nobody else was doing it here.

If it were you

Verdict: building another open-source Android input method isn't a good bet. In two months this slot has already filled with several projects with over a thousand stars, and input methods take a lot of engineering and a long time to tune their dictionaries. The part worth copying is its cold-start playbook. The author had only about 690 followers on Bilibili, no audience, and no big-company backing, and that approach works for others too. This sets it apart from many open-source projects that grew on the author's existing reputation.

Good fit: developers who already have, or plan to build, a working and fairly complete Chinese open-source utility, especially those making a replacement for an ad-heavy, always-online domestic app.

Bad fit: anyone hoping to collect stars with "yet another open-source input method," or anyone wanting to publish a video while the product is still a demo. CyIME's own argument cuts both ways: if you ship something rough, you become the kind of thing it criticizes.

How to do it:

1. Find a domestic app people are actively trashing. Beyond input methods, there are ad-laden file managers, photo galleries, readers, browsers, and cleanup tools. Search Bilibili for titles like "I gave up X app I used for X years"; anything with tens of thousands of views shows people are looking for alternatives.
2. Build on a mature open-source base and make it complete. CyIME uses Rime, the Wumeng dictionary, sherpa-onnx for speech, and upstream Xime, and spends its effort on being complete and detailed. Once the full version works offline right after install, you clear the bar that makes users willing to switch.
3. Appear on camera yourself and record a real-device demo of over 10 minutes. Make the title take a side and go against the trend, write a blunt bio, include the GitHub link, domestic mirrors (ModelScope, Netdisk), and a QQ group. A single video with 40k views can bring in hundreds of stars and tens of thousands of downloads.
4. After it takes off, ship daily for a week. Requests flooding the group and issues get fixed the same day when possible. CyIME kept adding over 60 stars a day on day 7 for exactly this reason.
5. Make sure domestic users can install it. Offer a lightweight and a full version, serve downloads via mirrors and in-app updates, and document troubleshooting for issues like installs failing on vivo / iQOO devices.

How it got users: the Bilibili video was the main driver, the QQ group caught people, and domestic download sites spread it on their own. X, HN, and Product Hunt were basically irrelevant in this category; CyIME has no discussion on any of them.

When this window closes: the "against vibe coding" and "I built a version that actually works" framing is still fresh. Once AI-generated similar projects are so common that everyone says it, the angle stops working. That's probably about six months away (inference). The sentiment around "replacing domestic ad-laden software" lasts longer and isn't bound by this window.

What would mean we misread it: if in a month or two CyIME's stars stall below a thousand, issues go unanswered, or the author's "taking a break" turns into a stop in updates, then one video brought only one-time traffic, and "thorough + daily releases" didn't keep people, so this playbook is worth less than it looks. Conversely, if someone builds a Chinese utility project of similar size purely through word of mouth on GitHub without a video, then the Bilibili step isn't essential.