50k downloads, 249 stars: a token-free "blue big fish" desktop pet
What it is A Windows/Linux desktop pet based on the "blue big fish," a whale-girl character derived from DeepSeek fan art. It includes a 4B local model, reads the text on your screen and pops up a line of commentary. Install by double-clicking the exe. Free and open source.
Why it took off It rode the October Bilibili "fish disaster" big-fish craze. Among similar tools that require your own API key and charge per sentence, it's the only one that uses no tokens, doesn't upload screen content, and anyone can install with one click. The UP主 Cznorth made a 580k-view comparison video around exactly that point.
CWhat you could do Stop building another big blue fish. Copy its template instead: a local small model + screen reading + a character persona + one-click install. Port it to macOS, which it doesn't have, or swap in the next AI meme as soon as it starts. Or ship it as an open-source framework where others fill in their own characters, and line up a Bilibili UP host to review it early.
- Data (updated daily) · Repo created 2026-10-03 · 276 stars · 7 days +117
GitHub · stars Data 2026-10-18 → 2026-11-08 (22 days, 22 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
In Oct 30, Bilibili UP host Cznorth posted a 71-second video titled "Blue Big Fat Fish Desktop Pet: 4B local model installed on your computer, no API Key, no tokens consumed (4060 real test)". It hit 90k+ plays in one day, with more bookmarks than likes (5,345 to 3,708), and reached 6k plays by Nov 583. The fish in the video was an open-source project created in Oct 3 by Bai-Cute, a GitHub account with a single repo and no social links. By Nov 6, its installer had been downloaded about 50k times, yet it had only 249 stars. This piece covers what it is, why it took off, where the stars came from, and whether it's worth copying if you want a project that gets stars too.
First, where the fish came from
Netizens showed DeepSeek its own logo, and the model, in its reasoning, identified itself as "a blue big fish." The community turned that into a character: a mermaid-ish whale girl with blue hair, a maid outfit, small whale fins by her ears, and a whale tail. In August, DeepSeek open-sourced DeepSeek Harness (DSH, 274k stars on GitHub), built around "everything is a plugin." Big-blue-fish desktop pets, QQ bots, and Bilibili-browsing plugins flooded in. Around Oct 22, a wave of DSH-driven whale-girl accounts swarmed Bilibili comment sections. It got called the "fish plague." Phoenix Tech 10-23 and Zhihu 10-24 both covered it, and the What's Worth Buying 11-02 article was titled "Day 8 of the fish plague, 4.61% of the comments are 'fish'."
The meme has a built-in joke: the big fish "eats rice," meaning it eats tokens. Raising a fish with DSH means every sentence it says costs you money.
What it is
A whale-girl desktop pet that lives on the edge of your Windows or Linux desktop. Double-click the exe to install. It bundles the Qwen 3.5-4B local model (Q4 quantized, about 2.7GB), llama.cpp, and all runtimes. It uses the GPU if you have one and runs on CPU if you don't. Every few minutes it uses the system's built-in OCR to read the text in your frontmost window, passes it to the local model, and shows one line in a cream speech bubble. If you're coding, she wags her tail and cheers you on. If you're chatting about dinner plans tomorrow, she gets "a little scheming" about the food. The README states that screen content is used only locally: nothing is uploaded or stored.
It also checks the weather hourly and reads you a tech news item from IT Home. It reminds you at mealtimes, sets off fireworks on holidays, and curtsies with a lifted skirt on startup. There are 106 transparent animation sets in total. The standard build can also connect to 21 command-line agents like Claude Code and Codex over ACP for long conversations. A separate Safe build removes that layer and keeps only the companion. It's free, Apache-2.0, with no paid tier.
Why this one, and not another big fish
The big-blue-fish desktop pet niche has been crowded since August. GitHub has dozens of repos on the theme, and Android download sites list 23+ apps (3DM roundup). The big ones:
| Project | Launched | stars (around 11-06) | How this one differs |
|---|---|---|---|
| PC2005-cloud/dsh-pet | 08-14 | 5,745, +762 in the last 7 days | Original animation, installs into DSH with one command, but can't talk on its own |
| MerZlin/dsh-pet-indesktop | 08-17 | 3,133 | dsh-pet moved into a standalone pet, no local model |
| Pal-AI-Lab/Coopanion | 09-23 | 2,940 | General AI desktop pet, needs your own API key, pay per use |
| QCYTSN/dsh-dafeiyu | 08-14 | 599 | Shows DSH Agent status, relies on cloud tokens |
| 1190fasheqi/dafeiyu-pet | 08-05 | 309 | Walks around and feeds, connects to DeepSeek API |
| QCYTSN/ds-local-pet | 08-06 | 45 | Also works offline, but no local LLM, no one-click install, didn't take off |
The character isn't theirs (it's a community fan creation), and neither is the animation. They use the 106 MIT-licensed animation sets from dsh-pet, upscaled with Real-ESRGAN and frame-interpolated with RIFE to 720p60, and the third-party notes say so clearly. What's unique to them is one combination: local model + screen reading + an exe an ordinary person can double-click to install.
That combination lands the joke. Cznorth's last video was about Coopanion, which he called "paying for every sentence." This one's title is "no API key, no tokens." He broke out the model and parameters and tested them on an RTX 4060 laptop: VRAM use 3.2GB, median reply time about 0.6 seconds (quoted by What's Worth Buying 10-31). The What's Worth Buying piece commented: "Reading the screen is exactly what cloud desktop pets wouldn't dare do. The local version turns it into the selling point."
The timing was lucky too. The v1.0.1 one-click installer shipped in Oct 18, four days before the fish plague peaked, and the build was already usable. That installer was later downloaded 16.6k times.
Where the stars came from: almost all Bilibili
Daily star count on GitHub (10-18 to 11-06):
| Date | star | What happened that day |
|---|---|---|
| 10-18 | 21 | v1.0.1 one-click installer released |
| 10-22 | 56 | Fish plague peaks |
| 10-30 | 105 | Cznorth's video published, v1.1.0 released the same day |
| 10-31 | 123 | +18 |
| 11-01 | 159 | +36, daily high |
| 11-02 | 184 | +25 |
| 11-06 | 249 | +144 over the last 7 days, 3.3x the first 7 days |
After the video, one v1.1.0 release alone got 20k downloads. Then came a round of reposts: a dedicated What's Worth Buying 10-31 feature, the Ciker software site 11-01, CSDN 11-04, and a few star-ranking aggregators. It doesn't show up on X, HN, Reddit, or Product Hunt. English-language communities aren't discussing it. Even the 36 Jike 11-06 fish plague roundup doesn't name it. Where the 16.6k downloads came from that week is unclear from the outside. Whether Cznorth and the author are connected isn't visible either. The video isn't labeled as marketing.
The number to remember: total installer downloads across all versions are about 51k (Windows about 48k, Linux .deb about 3,100, GitHub Releases, 11-06), against 249 stars. That's roughly 200 downloads for 1 star. The in-app auto-update also pulls the installer, so the real count of unique users is lower (estimate). Still, the scale tells us something: the people arriving were mostly not GitHub users. They were ordinary people who watched the video and came straight for the exe, and they don't click star.
What keeps people, and where it falls short
What keeps people is the details: weighted triggers for mealtimes, holidays, the lunar calendar, and bedtime. Inference stops automatically when unplugged to save power, and it hides during fullscreen games. You can pick a model from Hugging Face or China's ModelScope, with resumable downloads, or use it without downloading a model first. The author made 66 commits and 13 releases in a month, and the v0.5.0 release notes say "first time improving the experience based on community feedback." Outside developers have already sent PRs adding global weather and multi-monitor scaling, and someone has proposed a native macOS version. The Safe build accounts for about 30% of downloads, which suggests plenty of people do mind "letting the fish act."
The shortcomings are also plain. One commenter tried it and left two words: "very dumb." The 4B model only produces one line at a time, keeps no chat history, and remembers nothing, so anyone wanting a deep conversation has to plug in an external agent. Another comment put it bluntly: paying per sentence doesn't hurt in the moment. You only feel it when you cover the electricity and VRAM yourself.
If it were you
C. Worth learning from, not worth copying. Even the winner got only 249 stars in a month. Users are on Bilibili, not GitHub, so no amount of downloads turns into stars. The big-fish category has dozens of repos forking each other, and the biggest, dsh-pet, already has 5,700 stars. Growth came from one meme (the fish plague) plus one video with 580k views. Both are one-time events you can't replicate.
Who it fits: developers who can build Windows desktop software, package llama.cpp and a model into a one-click install, and already have some presence on Bilibili or in an anime or gaming community, so they can find UP hosts willing to review the product.
Who it doesn't fit: anyone hoping to earn stars by building "yet another big-fish desktop pet." The niche is full, the character and animation are public property, and whatever you build, someone can add next week. Also anyone who only writes code and knows no content creators. Stars for projects like this don't grow on GitHub.
How to do it: don't copy this fish, copy its template: "local small model + screen reading + character persona + one-click exe install, no API key." Three directions. First, a platform it doesn't cover. Only one community PR for a native macOS version exists, and it hasn't taken shape. Second, swap in the next AI meme or character that's just starting out, and have a finished product ready before the heat arrives (this fish shipped four days before the fish plague). Third, open-source the template itself as a framework so others can plug in their own characters and animations. That's a developer tool, and stars land on GitHub for tools like that. Use MIT-licensed community assets where you can, and put your effort into packaging and differentiating features. Also ship a "no agent actions" safe build.
How to get users: Find Bilibili creators who make AI tool reviews and give them a comparison angle. The hook here is "other tools charge per sentence, I spend zero tokens." Real benchmark numbers (VRAM, latency) make it easier to film. Put an animated GIF at the top of the README, ship the exe under GitHub Releases, and set up a domestic model mirror. To get stars, also ask people outside GitHub to "star it if you like it," and post it separately on developer hangouts like V2EX, Juejin and Hacker News.
When to move: a Mac version has to ship while the big fish is still hot, which means the next few weeks. The route of swapping in a new IP has no hard deadline, but you need to ship before the next meme breaks. A week late and you're the dozenth.
Signs you've misread this: (1) if this repo passed 1,000 stars by late November, or a Mac big-fish desktop pet gets several hundred stars in a month, the niche can still grow, and you should move it up a tier. (2) Conversely, if big-fish videos drop sharply on Bilibili in December and this repo's daily new stars fall back to single digits, the window for copying the template is closing too. (3) If a large project like Coopanion adds a local model option, "no tokens" stops being a unique selling point.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one PC2005-cloud/dsh-pet Open source 2026-08-14 5,745 stars, +762 in the last 7 days Original animation, DSH plugin installs with one command, no AI speech by itself. This project reuses its 106 animation sets, upscales and interpolates them, and adds a local LLM. MerZlin/dsh-pet-indesktop Open source 2026-08-17 3,133 star Cross-platform standalone port of dsh-pet with interaction, no local model, no screen reading. QCYTSN/dsh-dafeiyu Open source 2026-08-14 599 star Shows live DSH Agent status, depends on DSH and cloud tokens. 1190fasheqi/dafeiyu-pet Open source 2026-08-05 309 star Walking, feeding, and inner-monologue thoughts, connects to the DeepSeek API, which needs paid tokens. QCYTSN/ds-local-pet Open source 2026-08-06 45 star Also works offline, but no local LLM screen reading, no one-click install, didn't gain traction. Pal-AI-Lab/Coopanion Open source 2026-09-23 2,940 star General AI desktop pet with multiple personalities, needs your own API key and pays per use. Cznorth tested it last time, and this time uses "no tokens" as the contrast. VPet virtual desktop pet simulator Open source 2022-12 7,182 stars, listed on Steam Veteran desktop pet platform, not tied to the big-fish IP, no built-in local LLM. Android "Blue Big Fish Desktop Pet" series apps Android 2026-10 3DM roundup lists 23+ apps, data unclear Phone floating-window apps, mostly no AI DeepSeek Harness plugin ecosystem Big-company features 2026-08-13 deepseek-harness 274k stars No official desktop pet. User plugins require a DSH environment and tokens. - Where users come from
- Verified Bilibili review UP host video: Cznorth, "Blue Big Fish Desktop Pet: 4B local model on your PC, no API key, no tokens (4060 tested)" BV19WxP6YEME, posted 2026-10-30, 583k views; after release, v1.1.0 downloads 20k, stars 10-31 to 11-02, daily increase 18/36/25
- Verified Platform hot topic / meme wave (Bilibili "fish plague"): Phoenix Tech 10-23, Zhihu 10-24, What's Worth Buying 11-02, 36 Jike 11-06 covered whale-girl accounts swarming Bilibili; v1.0.1 got 16.6k downloads from 10-18 to 10-27
- Verified Media / content site reposts: What's Worth Buying 10-31 feature, CSDN 11-04, Ciker software site 11-01, pickfit 10-26, upstar.dev/ossradar star rankings
- Verified GitHub community / open-source contributions: external PRs from Tydecent, YamaFuso, zouyonghe and others (global weather, macOS version, etc.)
- Inferred X / English-language communities: x-radar search returned 0 results; none seen on HN/Reddit/PH
Why the Blue Big Fish Desktop Pet (Bai-Cute/big-blue-fish-desktop-pet) took off
Channel: GitHub | research date 2026-11-06 | data card: stars 249 / forks 16, last 7 days +144 (prior 7 days +43), 11-01 single-day +36
One line
Windows/Linux desktop pet: built-in Qwen3.5-4B local model (Q4, about 2.7GB, llama.cpp Vulkan). It uses system OCR to read text in the foreground window and periodically pops up a line in the voice of a whale-girl character; it can also read out the weather and cover IT Home news. The standard edition can connect to 21 CLI Agents via ACP (Claude Code, Codex, DeepSeek Harness, etc.) for long conversations. Free, Apache-2.0, no paid features.
Verdict
It rode the wave of the Chinese anime meme around the "blue big fat fish / fish disaster" and won on one angle in a crowded niche: no tokens burned. Other big-fat-fish pets at the time either only had animations and couldn't talk, or made users enter their own DeepSeek API key and paid per sentence. This one bundles a local 4B model with a one-click installer, plus the privacy pitch that "screen reading stays on your machine." What really drove its growth was a 71-second test video posted by Bilibili UP host Cznorth 2026-10-30 (currently 583k views). There's no evidence of paid promotion.
1. Key data
- Stars (our GitHub repo_day): 10-18 had 21 → 10-20 39 → 10-30 105 → 10-31 123 → 11-01 159 (+36, single-day high) → 11-06 249.
- Release download counts (GitHub API, 2026-11-06): Windows installer downloads across all versions total about 48k, Linux .deb about 3,100. By version: v1.0.1 (10-18) 16.6k, v1.0.5 (10-27) 6,655, v1.1.0 (10-30) 20.1k, v1.1.1 (11-05) 2,517. Note that in-app auto-updates also download the installer, so this number isn't unique users. But the download-to-star ratio is about 200:1, suggesting users are mostly not GitHub people; they're ordinary viewers who watched the video and went straight to download the exe.
- Timeline: the repo and GitHub account were created the same day (2026-10-03). The account has only this one repo and no social links. The v0.5.0 (10-16) release notes say "first time improving the experience based on community feedback." v1.0.0 (10-17) added MCP connection to Agents and a Linux build; v1.1.0 (10-30) switched to one-click ACP connection; v1.1.1 (11-05) upgraded animations to 720p60 and made weather location support worldwide. Iteration is very fast: 66 commits in one month.
- External contributors (issues/PRs): Tydecent (a string of PRs starting 10-17), YamaFuso (global weather, multi-screen scaling, persona), zouyonghe (submitted a native macOS PR). The community has already started adding features on its own.
2. Why it took off
1. Timing: "blue big fat fish" was one of the hottest Chinese internet AI memes in the autumn of 2026
- Origin: netizens showed DeepSeek's logo to DeepSeek, and the model said it was a "blue big fat fish." It was then remixed into a blue-haired, maid-outfit, whale-tailed "whale girl" (36 Kr 2026-11-06 "Big fat fish replaces Doubao, DeepSeek gets fan-created OCs and blows up," 36kr.com/p/4053852402504581; Gamersky 2026-08 reported that when V4-Flash launched, the character went viral, www.gamersky.com/news/202608/2190273.shtml).
- In August, DeepSeek Harness (DSH) was open-sourced, pitched as "everything is a plugin" (deepseek-ai/deepseek-harness, 274k stars, created 2026-08-13). From then on, big-fat-fish desktop pets, QQ bots, and Bilibili-scrolling plugins appeared in batches (same Kr article as 36).
- Around Oct 22, large numbers of whale-girl accounts driven by DSH flooded Bilibili comment sections, dubbed the "fish disaster" (iFeng Tech 10-23 tech.ifeng.com/c/8xSXpkNB7YG; Zhihu, "DeepSeek caused a fish disaster on Bilibili" 10-24; SMZDM 11-02 "Day eight of the fish disaster, 4.61% of the comments are 'fish'" post.smzdm.com/p/al395mlo/).
- This repo shipped v1.0.0,10-18 10-17 and v1.0.1 (with one-click install and GPU/CPU auto-detection), just days before the fish disaster broke out, so the product was already usable. The 16.6K downloads during v1.0.1's run mostly landed in the fish-disaster week (10-18 to 10-27).
2. Where users came from: Bilibili, with one review video as the main trigger
- Cznorth's Bilibili video "Blue Big Fish Desktop Pet: 4 Local Model Runs on Your PC, No API Key, No Tokens (4060 Hands-On Test)", BV19WxP6YEME, published 2026-10-30, currently 583K views (scraped from the page, www.bilibili.com/video/BV19WxP6YEME/). Smzdm 10-31 ran a feature piece saying it "pulled in over 90K views in a single day, with more favorites (5345) than likes (3708), the classic save-now-install-later pattern." It also points out that Cznorth's previous hit, Coopanion, required you to bring your own API key, and this one directly compares it with "every word costs money" (https://post.smzdm.com/p/apqr8l39/).
- The timing lines up: the video was posted on 10-30, v1.1.0 was released the same day, and it then picked up 20k downloads; stars rose 10-31 18, 11-01 rose 36 (single-day high), and 11-02 rose 25.
- Whether Cznorth has any connection to the author can't be confirmed. The video is labeled "contains AI-generated content" but not labeled as "promotional/marketing."
- Secondary spread: SMZDM (10-31), CSDN's "Github open-source AI desktop pet BigBlueFish blue big fat fish" (11-04, devpress.csdn.net/aibjcy/6aeb367840858c2d881888a2.html), the Cike Bianfeng software site (11-01), the pickfit install-tutorial site (10-26), plus star-ranking aggregators like upstar.dev and ossradar.
- No direct evidence found for where the downloads before the fish disaster came from (10-23 Bilibili has an install tutorial "Free to use entirely without consuming your own tokens" BV147nh69EX1,375 with some views, labeled "contains marketing information"; whether it's about this project is unverified).
- Channels I didn't see: searching x-radar for "Bai-Cute," "big-blue-fish," and "blue big fat fish desktop pet" each returned 0 results. No discussion in English-speaking communities; no traces on HN, Reddit, or Product Hunt; no signs of paid ads (it has no revenue, so it isn't buying paid acquisition).
3. Why people install it and keep it
- Zero cost: the big-fat-fish meme is itself about "eating for free = eating tokens," and raising a fish via DSH means paying for every sentence. A local 4B model turns that cost into a one-time "occupies 3.2GB of VRAM" (Cznorth tested it on a 4060 laptop; median reply time about 0.6 seconds, figures cited from SMZDM). It can also run on CPU without a graphics card.
- Low install barrier: double-click the exe; the runtime, .NET, and llama.cpp are all packed in. The model can come from Hugging Face or China's ModelScope mirror, with resumable downloads, and you can use it without downloading a model first. Most alternatives require installing DSH first, then running a command to install plugins.
- Animation quality: it directly uses the transparent animation from PC2005-cloud/dsh-pet (5,745 stars, MIT), the most popular in its category, with 106 clips upscaled to 1280×720 and 60 frames using Real-ESRGAN + RIFE (documented in code/THIRD-PARTY.md), plus weighted triggers for mealtimes, holidays, the lunar calendar, and sleep.
- Screen reading + privacy: it can comment on what you're looking at, and what it reads is not uploaded. Cloud pets find this very hard to do (SMZDM's words: "screen reading is exactly what cloud pets dare not even imagine; the local version turned it into a selling point").
- Weakness: commenters say it's "dumb." A 4B model outputs one sentence at a time with no memory, so long conversations require connecting an external ACP Agent.
III. Competitors
The big-fat-fish desktop pet niche is very crowded. GitHub has dozens of repos on the same theme, and Android download sites list more than 23 "blue big fat fish desktop pet" apps (3DM feature shouyou.3dmgame.com/zt/deepseeklsdfyzc/).
| Product | Where | Launched | Scale | How it differs from this project |
|---|---|---|---|---|
| PC2005-cloud/dsh-pet | Open source (DSH plugin) | 2026-08-14 | 5,745 stars, +762 in the last 7 days | Original animation (this project uses its assets); one-line command to install into DSH; no AI speech of its own |
| MerZlin/dsh-pet-indesktop | Open source | 2026-08-17 | 3,133 star | Ported dsh-pet into a standalone Win/Linux/Mac desktop pet with interactions added; no local model |
| QCYTSN/dsh-dafeiyu | Open source (DSH) | 2026-08-14 | 599 star | Shows real-time status of DSH Agents; depends on DSH and cloud tokens |
| 1190fasheqi/dafeiyu-pet | Open source | 2026-08-05 | 309 star | Can walk and be fed, shows "chain-of-thought inner voice"; connects to DeepSeek API |
| Sutera-Diffusus/dsh-whale-musume | Open source (DSH) | 2026-08-15 | 205 star | Gamified: achievements, mini-games |
| QCYTSN/ds-local-pet | Open source | 2026-08-06 | 45 star | Also focuses on offline use (PySide6), but has no local LLM screen reading and no one-click installer; didn't take off |
| RYANFFY/deepseek-dafeiyu-pet, whiteball-shiro/blue-whale-agent, etc. | Open source | 2026-09 | A few to a few dozen stars | Mainly shows DeepSeek balance |
| Pal-AI-Lab/Coopanion | Open source | 2026-09-23 | 2,940 star | General-purpose AI desktop pet with multiple personalities; requires your own API key and pay-as-you-go (Cznorth's previous episode tested this one) |
| VPet virtual desktop pet simulator | Open source / Steam | 2022-12 | 7,182 star | Long-running desktop pet platform, not tied to the big-fat-fish IP, and no built-in local LLM |
| Android "Blue Big Fish Desktop Pet" series apps | Android download sites | 2026-10 | More than 23 apps, data unclear | Phone floating-window pets, mostly without AI; Tony-Chen-Stark/dsh-pet-android (113 stars) is an Android port of dsh-pet |
| DSH itself + plugin marketplace | big tech / open source | 2026-08 | 274k stars | Officially no desktop pet; the ecosystem fills the gap with user plugins |
Bottom line: neither the character nor the animation is its own (the animation comes from dsh-pet; the character is a community-shared OC). What it uniquely has in this niche is the combination of "local LLM + OCR screen reading + one-click exe install." The earlier offline route, ds-local-pet, had no LLM and never took off.
4. Differentiation, and whether it's the reason it took off
- Differentiation: a local AI desktop pet with no API key, no tokens burned, and screen reading that isn't uploaded, packaged as a finished product an ordinary person can install by double-clicking; later it added ACP to connect 21 Agents, leaving "can chat and get work done" for those willing to tinker.
- Is it the reason? Yes. The Bilibili video that pulled in traffic is titled after this very differentiator ("no API key, no tokens burned"), and it was shot as a comparison against Coopanion, which charges per sentence. Animation quality and holiday details determine whether people stay after installing, but they didn't drive acquisition.
5. Visibility and copycats
- Visibility: circulating in small circles. It has been noticed in the Chinese-language circles on Bilibili, 什么值得买 and CSDN (one video at 580k views, plus a dedicated article). But the GitHub repo has only 249 stars, there is 0 discussion on X and in the English-speaking community, and 36 氪's fish-disaster roundup doesn't name it. No revenue, and no plans to charge (free and open source, no donations or paid tier).
- Copycats: so far I haven't seen a clear project copying the "local-model big-fat-fish desktop pet." The whole big-fat-fish pet category has always been people forking and porting each other's work (dsh-pet → dsh-pet-indesktop → Android versions → shenliuyu18/dafeiyu-pet, etc.), and this project itself started from dsh-pet's assets. Cloud-based pets like Coopanion will probably add a local-model option eventually.
6. What others can borrow
- When riding a meme IP, hit the "cost" pain point: the IP already carries the "eats tokens" joke, so a local model becomes a natural opposing selling point, and review creators are willing to use it for comparison videos.
- Finished-product form for non-developers: one-click exe, a China model mirror, resumable downloads, usable even without a model. Downloads and stars differ by 200:1, showing the real users are outside GitHub.
- Reuse MIT-licensed community assets, then add technical enhancements (upscaling + frame interpolation), and put your effort into differentiating features.
- Safety tiers: ship both a standard edition and a Safe edition (ACP removed, Agents can't take actions), easing concerns for ordinary users. The Safe edition accounts for about 30% of downloads.
- This template (local small model + OCR + character persona + one-click install) can swap IPs (other AI waifus, game characters) and platforms (someone has already asked for a Mac PR).
VII. What I'm still unsure about
- No direct evidence found for where the 16.6k downloads from the fish-disaster week (10-18 to 10-27) came from.
- Whether Cznorth has any relationship with the author or any collaboration: unverified.
- Unknown how much of the downloads were in-app auto-updates, how many unique users there are, or the DAU and retention.
- The fish disaster is meme-driven, so how long the heat lasts is uncertain. The IP itself isn't officially owned (DeepSeek hasn't claimed it), and the risk of copyright issues or official action down the line is unknown.
- Other Bilibili videos mentioning it (Douyin, Xiaohongshu) haven't been systematically checked.