# Stop building a "mini Jev", help Jev save money instead: how hunch got 400 GitHub stars with one tweet and a few browser games · open source project

> Source: https://stringers.io/en/github/gh_undrgrnd-labs__hunch.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> It landed right as Jev's decision-model buzz and the EmbeddingGemma 2 launch were both happening, and the post tied the two together. The author of Deno posted a comparison GIF with a small game you can play in the browser. A week later the Chinese AI account AIGCLINK reposted it, which brought in another wave.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

undrgrnd-labs/hunch · undrgrnd-labs

# Stop building a "mini Jev", help Jev save money instead: how hunch got 400 stars with one tweet and a few browser games

**What it is** An open source "reflex layer" that sits in front of a judge model. It uses EmbeddingGemma 2 vectors and a small classifier to learn the judge's answers. Once it's accurate enough, it answers on CPU by itself, and hands anything it's unsure about back to the judge.

**Why it took off** It landed right as the decision-model wave kicked off by Jev and the fresh release of EmbeddingGemma 2 were both peaking, and it connected the two. A developer who had built Deno posted a tweet with a comparison GIF, and a week later the Chinese AI account AIGCLINK reposted it, bringing in another wave of attention.

**C** **How you could follow it** Don't build another distillation proxy. Copy the playbook instead: build a middle layer that saves money next to a new paradigm that's already hot, tie it to a freshly released open model from a big lab, put together a screen-recorded web demo, and after launch pitch it to Chinese AI news accounts for reposts.

C grade: hunch and the earlier stuntd already occupy this spot, so a third entry would be hard to notice. With the same idea, the lesser-known stuntd got only 74 stars. The few hundred extra stars hunch picked up came mostly from its author's 10,000+ followers and the Chinese account reposts, and most followers won't get those. What you can learn is the playbook, but you'd have to bet on it again when the next new model drops. What's being whispered in small circles

Data (updated daily) · Repo created 2026-10-21 · 490 stars · 7 days +197

GitHub · stars

Data 2026-10-30 → 2026-11-08 (10 days, 10 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

hunch is an open source "reflex layer". It sits in front of decision models like Jev and Clef, or LLM judge models. It watches the teacher answer questions and learns from them. Once it's accurate enough, it makes its own calls, producing a verdict in about 20 milliseconds on an ordinary CPU, and it hands anything uncertain back to the teacher. It was created in Oct 21. In Nov it went from 6 to 413 stars (GitHub daily star count), and it jumped only twice: once when the author tweeted about it, and once when a Chinese AI account reposted it a week later. The idea itself isn't unique. A nearly identical project had appeared a month earlier and only reached 74 stars. So this piece is about where those extra few hundred stars came from, and which parts you can copy and which you can't.

### One tweet, star number 17

On 2026 Oct 28 18:07 (UTC), Matt Mastracci (@mmastrac) posted a tweet on X:

> Meet hunch! Jev-like decisions in ~20ms without even needing a GPU, learning live from a teacher model. Built on @GoogleGemma's EmbeddingGemma 2, it can works with any teacher model like @Cloudflare's Clef.

The tweet included a comparison GIF. The same batch of agent tool calls was judged by hunch, Clef, and Galileo Luna on whether the call had failed. hunch was the fastest and cheapest, and Luna was the slowest and most expensive. In the next tweet he explained the background: he built it while working at undrgrnd, a stealth startup doing continual learning. "We needed to scan a lot of private data, so we built hunch. Then we realized it was a really useful building block, so we open sourced it. Give us a ⭐️."

Someone in the replies wrote: "Marking this, I'm star number 17. Feels like this is going to blow up. I'll try swapping the Jev in my own system for it later."

The tweet later got 411 likes and 12.4k views (per jev.guide's stats). Mastracci has 11.2k followers on X and has worked on Deno, StumbleUpon, and FullStory, and was once a CTO. This detail will come up again below.

### What it is

Teams building agents share the same headache. Every run of an agent spits out hundreds of lines of text, code, and tool calls, and you need to check in real time: did this step's call fail, is there a prompt injection in this output, does it leak personal information? Those same few questions have to be asked for every record. The usual approach is to call a judge model each time. Even the fastest takes hundreds of milliseconds, and asking a lot of questions makes the bill ugly.

hunch's idea is that "most of these questions are simple most of the time." It sits in front of the judge model (which it calls the "teacher"):

1. **Learn** : for each question, a small classifier (the "student") is trained on vectors computed by Google's EmbeddingGemma 2, using the teacher's answers as labels.

2. **Graduate** : cross-validation shows the student agrees with the teacher often enough, so the student starts answering on its own.

3. **Spot-check** : even after graduation, some answers are still sent back to the teacher for review. Inputs that are too far from anything it has learned go straight to the teacher.

Its tagline is "the judge model is still in the loop, it's just called less often." According to the README, judging a tool-call result of about 5 tokens on an Apple M190 Pro takes a median of 22 milliseconds. The underlying EmbeddingGemma 2 runs on eg2cpu, a Rust CPU engine the author wrote, so no GPU is needed. It only supports three kinds of questions with fixed answers: yes/no, pick one from several options, and rate on a few levels. The README is also candid about the weak spot: hunch is fast because it doesn't think. Questions like "how severe is this failure and who should fix it" that require multi-step reasoning still go to the teacher.

The repo also includes a `hunch.live` library that can learn text and images while you use it in the browser. The author built four browser games on top of it at tryhunch.dev: Pictionary, where you draw and it guesses; rock-paper-scissors, where you first teach it your hand gestures and then play against it; an emoji guessing game; and a camera Q&A where a vision model answers first and hunch takes over once it has learned.

### The two waves it caught

- **The decision-model craze.** TypeSafe released Jev in Sep 15 (early access, weights not released). Within a month a wave of related projects appeared on GitHub: laya, a zero-shot decision model with 35k stars; jaredpalmer/kev, which fine-tunes Qwen so you can "train your own Jev" with 9,800 stars; plus awesome-jev and AnyJev. Cloudflare open sourced Clef and Clef-flash in Oct 1.

- **EmbeddingGemma 2 had just dropped.** Google released it in Oct 6, and the community was looking for uses beyond retrieval. Some were already using it for image classification and ad detection on CPU.

Everyone else was building "smaller, faster Jev". hunch was building "call Jev fewer times", and it also gave EmbeddingGemma 2 a new use case. People on both sides had reasons to share it.

### Where the stars came from: two spikes

| Date | What happened | star |
|---|---|---|
| 10-28 | The author posted the launch tweet and tagged Google Gemma and Cloudflare. | Starting point |
| 10-29 | The author released the browser Pictionary: "no server, data never leaves your machine." Jev community aggregators (the jev.guide showcase, jevainews, jevhunt, and others) and two Korean tech blogs picked it up. | +77 that day (Trendshift) |
| 10-30 to 11-03 | No new spread, single-digit new stars per day. | 269 → 294 |
| 11-04 | An English post said "people are finding new uses for EmbeddingGemma 2 beyond retrieval... a community project". The author didn't track it down. | +34(Trendshift) |
| 11-05 | AIGCLINK (@aigclink) posted a long Chinese tweet with a video at 07:04 UTC: "Someone just found a new trick with Google's EmbeddingGemma 2... let the big model be the teacher, and let a local small model learn while it's being used." It got 38.3k views, 496 likes, and 618 bookmarks. AITOP and AGI Hunt reposted it. | 316 |
| 11-06 | The aftermath of the Chinese tweet from the day before. | 413(+97,fork 26 → 36) |

Star counts come from GitHub's daily records. Trendshift's day-by-day star gains don't match its absolute numbers (both Trendshift and the tryhunch.dev homepage show over 470), so I only used them to compare shapes, and the two shapes match. It never made GitHub's trending page, and I found no Show HN post, no Reddit thread, and no paid promotion.

Each of the two spikes lines up with one person: the first was the author, and the second was AIGCLINK. 618 bookmarks against 496 likes is typical of "save it for later", so the 11-06 97 stars can largely be credited to that tweet (inference).

### The same idea, 74 stars

hunch wasn't the first to come up with "learn from the teacher first, then take over once you've learned enough."

| Project | Launched | star | How it differs from hunch |
|---|---|---|---|
| bladedevoff/stuntd | 09-23 | 74 | Almost the same idea: a local agent learns from Jev or OpenAI's answers and takes over once it's learned. Built on laya, no browser demo, and the author had no following. |
| adrida/tracer | 03-29 | 1,036 | Offline distillation of LLM traces into a small model. hunch learns online and keeps spot-checking after graduation. |
| RouteLLM / semantic-router | 2023–2024 | 5,700 / 4,000 | Routing between big and small models, or routing by vector. Doesn't keep learning from a teacher. |
| huggingface/setfit | 2022 | 2,900 | Few-shot classification that needs manual labels. hunch has the teacher act as the labeler. |

Numbers from the GitHub page. stuntd came a month earlier with the same idea, and the difference came down to three things. hunch's base is an open model just released by a big lab, so its launch tweet could tag the official account. It had small games you can play directly, which are easy to screen-record (the video in AIGCLINK's tweet was made from exactly this material). And the author brought over 10,000 followers. Differentiation is what makes a project take off, but the few hundred extra stars mostly came from name recognition, the demos, and the Chinese account reposts.

Also look at how big it is: 16 days and 400 stars, which is small next to the Jev wave, and two orders of magnitude below laya's 35k. It spread within a small circle and never broke out.

### If it were you

**In one sentence** : rebuilding another hunch isn't worth it, but copying its playbook is. **Build a middle layer that saves money next to a new paradigm that's already hot, tie it to a freshly released open model from a big lab, and ship a demo you can play in the browser.** But go in with realistic expectations: without the author's 10,000+ followers, the same thing will probably get stuntd's 74 stars, not hunch's 400.

**Who it suits** : developers who can write both model inference and front-end demos, who watch for new model releases and can put together "new model + new use case + a playable web page" within a week or two, and who ideally already have some followers on X or can talk to Chinese AI news accounts.

**Who it doesn't suit** : anyone who wants to build yet another "distillation proxy in front of a judge model" (hunch and stuntd already each took one approach, so a third is hard to notice); anyone with no following who won't go find accounts to repost them; and anyone hoping to make money from it (for projects like this, the payoff is reputation and credentials. undrgrnd open sourcing it looks more like building a technical brand and hiring).

**How to do it** , from most to least reliable:

1. **Build a "savings layer" around something more expensive** : hunch saves on judge-model calls. The same "learn, graduate, spot-check" structure can be applied to other calls that are expensive and repetitive: vision LLMs looking at screenshots, OCR, reranking, content moderation. Pick whichever is hottest and most expensive right now, and use a freshly released small open model as the student (inference).

2. **Build it in a different ecosystem** : hunch needs Rust and uv to run. A pure TypeScript version that runs in the browser or Node with WebGPU, or a version that defaults to Qwen or DeepSeek as the teacher with clear Chinese docs, would serve people hunch doesn't reach (inference).

3. **Do what it hasn't done** : the multimodal section of the proxy layer says "coming soon", and it isn't hooked up to trace platforms like Langfuse or LangSmith yet. A plugin that lets you "install it and take over the repeated judgments in your trace platform" would be instantly clear to users (inference).

**How to get users** : follow hunch's order. In the launch tweet, state clearly "X times faster, Y times cheaper", attach a comparison GIF, and tag the official account of the model you use. The next day, put up a zero-install web mini-game designed for others to screen-record. Then submit it to the paradigm's awesome lists and showcase sites. A week later, approach Chinese AI news accounts (the AIGCLINK type), where one tweet can bring in hundreds of stars. hunch never posted on Show HN, so you can fill that gap.

**Timing** : this playbook works in the few weeks after a new model comes out, while everyone is looking for new uses. EmbeddingGemma 2 was released in Oct 6, and interest fades after that. A more realistic move is to wait for the next open model from a big lab and launch within two or three weeks of its release (inference).

**Where the math goes wrong** : ① hunch gained 1,000 stars over a month on its own, without new reposts. That suggests the idea itself pulls people in, not just the author's name, so it's worth seriously building a 2 and a 3; ② stuntd and other lesser-known authors who built this approach still reached a few hundred stars, so the playbook is copyable by ordinary people; ③ conversely, if Jev released its weights or cut prices sharply, or if models like Clef got fast and cheap enough on their own, a middle layer that "saves money" would be far less interesting.

---

Sources and attribution: the GitHub repo undrgrnd-labs/hunch page and README (11-06); GitHub daily star counts (10-30 to 11-06); Trendshift's day-by-day star gains; Matt Mastracci's launch post and replies on X (10-28, 10-29), with likes and views from jev.guide; AIGCLINK's Chinese post on X (11-05); the tryhunch.dev homepage; GitHub pages of each competitor; and the Google and Cloudflare launch blogs. I couldn't identify the author of the English post at 11-04. I can't tell how many people actually use it or whether the stars are inflated. The link between the retweets and the star jumps is inferred.

[GitHub repo ↗](https://github.com/undrgrnd-labs/hunch)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| TypeSafe Jev | Other | 2026-09-15 | The source of this wave of decision-model hype: a commercial API, in early access. | Jev is the teacher model, and hunch sits in front of it to cut down calls. |
| Cloudflare Clef / Clef-flash | Big-company features | 2026-10-01 | Hosted on Workers AI, open weights on HF. | Also a teacher-class model, but it needs a GPU. hunch runs on CPU only and learns online. |
| NandhaKishorM/laya | Open source | 2026-09-18 | 35K stars | A zero-shot decision model. hunch learns from the teacher per problem. |
| Jev clones such as jaredpalmer/kev, fine-tuned on Qwen. | Open source | 2026-09 | 1000~9800 stars | These require training and a GPU. hunch uses no GPU and learns while it runs. |
| bladedevoff/stuntd | Open source | 2026-09-23 | 74 stars | Nearly the same idea (a local proxy that learns Jev's answers and then takes over), built on Laya. hunch uses EmbeddingGemma 2 plus a Rust CPU engine, supports multimodal input, ships a browser demo, and its author has a name in the field. |
| adrida/tracer | Open source | 2026-03-29 | 1036 stars | Distills traces offline into a System 1 model. hunch is an online proxy that keeps spot-checking after it graduates. |
| RouteLLM / semantic-router | Open source | 2023-2024 | 5700 / 4000 stars | Does routing or vector routing, and doesn't keep learning from the teacher. |
| huggingface/setfit | Open source | 2022 | 2900 stars | Needs human labels. hunch has the teacher model label data automatically. |

Where users come from

- **Verified** **The author's own launch post on X** : x.com/mmastrac/status/2115505657231384683,10-28,411, 12.4k views per jev.guide's count. The author has 11.2k followers and built Deno.

- **Verified** **Browser mini-game demos (Pictionary and others on tryhunch.dev)** : the author's post at 10-29, x.com/mmastrac/status/2115886824078131521. Trendshift recorded 10-29 +77.

- **Verified** **Jev ecosystem aggregator sites and Korean blogs** : jev.guide, jevainews.com, jev-x.com, jevhunt.com, enghyu.com, roomedia.tistory.com.

- **Verified** **Chinese AI news accounts reposting it** : AIGCLINK 11-05 post x.com/aigclink/status/211823790671652B views/618 bookmarks; AITOP and AGI Hunt repost; matches the surge in 11-05/06

- Inferred **HN / Reddit** : Algolia and Reddit searches turned up nothing.

## Research report on undrgrnd-labs/hunch

> In one line: hunch is a "System One proxy." It sits in front of Jev, Clef, or LLM judge models, uses vectors from Google's EmbeddingGemma 2 plus a small classifier, and learns the teacher's answers. Once it's accurate enough, it answers on its own (about 20ms on CPU). Anything it can't answer still goes to the teacher. Repo: [github.com/undrgrnd-labs/hunch](https://github.com/undrgrnd-labs/hunch), homepage: [tryhunch.dev](https://tryhunch.dev). Repo created 2026-10-21, released 10-28.

### 1. Data

| Date | star | fork | Source |
|---|---|---|---|
| 2026-10-30 | 269 | 23 | our GitHub repo_day |
| 2026-11-01 | 293 | 26 | Same as above |
| 2026-11-04 | 296 | 26 | Same as above |
| 2026-11-05 | 316 | 26 | Same as above |
| 2026-11-06 | 413 | 36 | Same as above (gain_1d 97). |

- Trendshift's daily star gains (https://trendshift.io/repositories/324652): 10-29 +77, 10-30 +14, then single digits for the next few days. 11-04 +34, 11-05 +102, 11-06 +44. The current star count shown on the page is 474. It doesn't match our database's absolute numbers, but the shape matches: **two spikes** , the first at the 10-28/29 launch and the second from 11-04 to 06.

- It never made GitHub trending (trending_days_any=0). It hadn't been up for 60 days anyway, so it couldn't qualify.

- 2 contributors and 9 merged PRs (per Trendshift).

### 2. Where the growth came from (by time)

1. **The author's own launch post on X (10-28, the first spike)** . The author is Matt Mastracci (@mmastrac). He has 11.2k followers on X, built Deno, StumbleUpon, and FullStory, and was once a CTO (from his GitHub and LinkedIn). The launch post, "Meet hunch! Jev-like decisions in ~20ms without even needing a GPU…", is at [x.com/mmastrac/status/2115505657231384683](https://x.com/mmastrac/status/2115505657231384683). According to jev.guide it has **411 likes and 12.4k views** . The post tagged @GoogleGemma and @Cloudflare and included a side-by-side GIF. One commenter said, "I'm star number 17, this feels like it's going to blow up."

2. **A browser demo you could play with went up the next day (10-29)** : you could play Pictionary with hunch, "no server, data never leaves your machine" (https://x.com/mmastrac/status/2115886824078131521). tryhunch.dev later added rock-paper-scissors, emoji charades, and webcam Q&A. Trendshift recorded +77 that day.

3. **Auto-indexed by Jev ecosystem aggregators** : sites such as jev.guide's showcase, jevainews.com, jev-x.com, jevhunt.com, and the awesome-jev list all picked it up. The Korean tech blogs enghyu.com and roomedia.tistory.com also wrote about it around 10-29. This is long-tail exposure.

4. **A Chinese mega-account drove the second spike (11-05)** . AIGCLINK (@aigclink) posted a long Chinese thread with a video at 11-05 07:04 UTC: [x.com/aigclink/status/2118237906716516496](https://x.com/aigclink/status/2118237906716516496), which got **38.3k views, 496 likes, and 618 bookmarks** . A GitHub link in the replies got 1800 views. The same day AITOP/AI热报 and AGI Hunt reposted it, and AIGCLINK's Notion page archived it. The Korean Threads account @choi.openai also shared it (3200 views). The +34 in 11-04 came from an English post Trendshift logged ("People are finding really creative ways to use EmbeddingGemma 2 beyond search… community project"). I couldn't identify its author.

5. **Channels I couldn't find** : no Show HN (Algolia finds nothing), no Reddit threads, no paid promotion.

Takeaway: **the author's personal reach and a launch post set the baseline, then a Chinese AI news account reposted it a week later** . The +97 on 11-06 is almost certainly the effect of AIGCLINK's post. The timing fits, and its bookmark count is high.

### 3. Timing: which wave it rode

- **The "System One / decision model" wave** : TypeSafe AI launched Jev in 2026-09-15 (early access, weights not public). Over the next month a batch of Jev-related projects appeared on GitHub: laya (35k stars), jaredpalmer/kev (9800), AnyJev (1600), awesome-jev, and others (our github table). Cloudflare open-sourced Clef and Clef-flash in 10-01 (https://blog.cloudflare.com/clef-decision-models).

- **EmbeddingGemma 2 had just launched** : Google released it in 2026-10-06 (https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/), and the community was looking for uses beyond retrieval. People were already using it for CPU image classification, ad detection, and WebGPU classification in the browser. hunch connected the two trends: it uses EmbeddingGemma 2 vectors to "cheaply replace Jev / Clef."

- The pain point is concrete: judgments on agent traces, tool calls, and prompt-injection detection come in high volume and keep repeating, and calling a judge model every time is slow and expensive.

### 4. Competitors

| Name | Where | Launched | Scale | How hunch differs |
|---|---|---|---|---|
| TypeSafe Jev | Commercial APIs | 2026-09-15 | The source that kicked off the wave: early access. | Jev is a teacher-class decision model that needs a GPU and charges per call. hunch sits in front of it and cuts down the number of calls. |
| Cloudflare Clef / Clef-flash | Big-company feature + open weights. | 2026-10-01 | Hosted on Workers AI, open on HF. | Same as above. Clef is also a teacher inside hunch. |
| NandhaKishorM/laya | Open source | 2026-09-18 | 35K stars | A zero-shot, non-autoregressive decision model. hunch doesn't do zero-shot. It distills from the teacher for each specific problem. |
| jaredpalmer/kev, Mapika/decider, firelex/jeff, StartLux-Decision, and others. | Open source | 2026-09 | 1000~9800 stars | All are Qwen fine-tunes that "train your own Jev" and need a GPU. hunch runs on CPU only and learns online. |
| bladedevoff/stuntd | Open source | 2026-09-23 | 74 stars | **Nearly the same idea** : a local proxy that learns the answers from Jev/OpenAI and takes over once it has learned them. The difference: stuntd is built on Laya, while hunch uses EmbeddingGemma 2 plus its own Rust CPU engine. It supports multimodal input, ships a browser demo, and its author has a name in the field. |
| adrida/tracer | Open source | 2026-03-29 | 1036 stars | Distills LLM traces into a System 1 model, mostly offline training. hunch is an online proxy that keeps spot-checking after it graduates. |
| lm-sys/RouteLLM, aurelio-labs/semantic-router. | Open source | 2023~2024 | 5700 / 4000 stars | One routes between large and small models, the other does vector routing. Neither keeps learning from a teacher. |
| huggingface/setfit | Open source | 2022 | 2900 stars | Few-shot vector classifiers that need human labels. hunch has the teacher model label data automatically. |
| Judge models such as Galileo Luna. | Commercial | — | — | hunch's demo uses it as the slow, expensive baseline for comparison. |

### 5. Differentiation

- **Different shape** : others are building "smaller, faster decision models." hunch builds a teacher-student proxy layer: learn, graduate, spot-check, and fall back to the teacher when unsure. The pitch is "the judge model stays in the loop, it just gets called less."

- **Engineering** : EmbeddingGemma 2 runs on the custom Rust CPU engine eg2cpu, so no GPU is needed. The same setup handles both text and images. It exposes a Jev-style systemone interface that you can drop into an existing system.

- **A demo you can play with right away** : Pictionary, the webcam, and gestures all run in the browser. This matters a lot for spread. The video in AIGCLINK's post came from these clips.

- The differentiation was **necessary but not the main reason** it took off. stuntd, which has nearly the same idea, launched a month earlier and has only 74 stars. hunch reached 400+ mostly because of the author's reputation, a demo that is easy to film and share, and the repost from a Chinese mega-account.

### 6. Visibility and copycats

- Visibility: **it spread within small circles** . In English, the main sources were the author's own post (12k views) and the Jev ecosystem aggregator sites. In Chinese, it was one AIGCLINK post with 38k views and a few news accounts. No mainstream media coverage, and nothing on HN.

- Copycats: none seen yet. tareqmahmood/hunch is just a fork. stuntd came earlier with the same idea, so it doesn't count as a copycat.

- No revenue involved. undrgrnd is a stealth startup working on continual learning. Open-sourcing hunch looks more like building technical credibility and recruiting than making money directly.

### 7. What to borrow

1. In a hot new paradigm, a "middle layer that cuts its cost" is easier to notice than yet another model of the same kind.

2. Tying it to a freshly released big-company open model (EmbeddingGemma 2), tagging the official accounts in the launch post, and getting a second round of spread from "new uses for the official model."

3. Building a zero-install browser mini-game demo that news accounts can screen-record and repost easily.

4. Chinese AI news accounts (like AIGCLINK) drove a clear lift for this open-source library: one post brought in roughly 100 stars.

### 8. Uncertain

- Who posted the original English tweet 11-04 is unclear (maybe someone tied to Google Gemma or HF). I couldn't find out.

- The absolute star counts from Trendshift and our own library don't match (474 vs 413), so I only used them to see the shape of the curve.

- I couldn't find real usage data (PyPI downloads, whether anyone runs it in production). The GitHub API hit a rate limit, so I couldn't pull the stargazer list and judge whether stars were bought (our library's suspect field is empty).

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.