1,600 stars, 0 downloads: an open-source library that turns AI Dungeon into video, and its fork every 20 minutes

Murea-InWorld is an open-source interactive audio-visual world model: you type "what happens next," and it films the story as a continuous short clip with sound. In 24 days it went from 25 stars to over 1,600, which looks like a project gaining momentum. Taken apart, the real exposure is one set of articles from a Chinese AI outlet, and the rest of the growth has a clearly machine-like rhythm. Below: what it is, how the numbers came about, and whether this direction is worth building for you.

Five forks from the evening of Nov 6

On Nov 6 of 2026, the five newest forks of InWorld were created at Beijing time 16:44, 17:26, 18:05, 18:45, and 19:05, by accounts named ayasma37, zerocearlo, jjohnnettenduara, yxemg, and atrothasantomaim, and the forked repos were renamed hzjlwiaow, tawafqclqgx, rnpujnksje, qcireoghnj, and jujtxcbncq (GitHub, 11-06).

This isn't a coincidence. Looking back, 259 forks in total, 229 of them were created just after the top of the hour, at :02, :22 or :42, within a few seconds of each other. Starting Oct 25 00:02:09, that's 19 to 23 a day, without a single day off. The accounts I spot-checked each have 0 followers, 0 following, and 1 to 5 repos. The same accounts also forked other repos, like one project with 23 stars and 251 forks. This is a pool of accounts shared by several parties.

What it is

Murea Labs is a GitHub org created in Oct 2026 12, with just this one repo. Judging from the Chinese write-ups and partners, it's most likely a China-based team (inference). InWorld has two parts:

• Murea-Mind: a "world state" agent that acts as screenwriter and director. It keeps track of characters, items, locations, and plot progress, reads your input and the last frame of the previous clip, plans the next event, and after generation reviews the footage and writes the actual changes back to the "ledger." It calls Gemini, OpenAI-compatible APIs, or Alibaba's DashScope underneath.
• Murea-Fast: a generation service built on Lightricks' open-source LTX video model, where characters, cats, props, and scenes each get a reference image to lock their look, and the video continues segment by segment, with picture and sound produced together.

It supports rollback and branching, and 2 to 4 people can co-write a story in one room. The official site's first screen is a 2-minute 19-second demo video with the line "Create a world, choose your next move, and watch the story unfold with video and sound." In short, it "films" the kind of text adventure found in AI Dungeon.

Two things to watch: the top-scoring flagship Murea-Pro is not open source; what's open is the lightweight Murea-Fast. And according to the docs, Fast's weights need 4 H20 GPUs to run, and there is no online demo on the web.

Where the stars came from

A set of media articles, with real exposure, but it didn't move the numbers. On Oct 18, Synced published "Top of Two Leaderboards: What They Want Is Not the Next World Model," covering how Murea-Pro scored 81.1 to take first place on Meituan LongCat's interactive video benchmark WBench, and how Murea-Fast ranked first among open-source models on OpenS2V-Eval. Tencent News and Sohu reposted it. On Oct 25 a video followed (the YouTube version had 92 views), and on Oct 27 Synced's X account posted an English tweet tagged "#PapersAccepted," the name of its paper-submission column. The articles read throughout like promotional copy, like a team submission or a sponsored piece (whether it was paid is unverified). In the week after publication, stars only rose from about 25 to 29.

Then the curve suddenly flattened. From Oct 25 the count went from 29 stars, with 26 on 163, and afterward gained +77 to +153 a day, most days holding steady between 120 and 135 (GitHub daily data, 10-21 to 11-06). On Nov 6 it was 1,559 stars, and a live check that afternoon showed 1,693 stars and 251 forks. Normal media exposure spikes and then falls off; it doesn't stay at nearly the same daily amount for over ten days.

Three numbers that don't add up. Behind the 1,600+ stars: watchers are 0, open issues are 0, and the weight downloads on Hugging Face are 0 (Hugging Face, 11-06). The last code commit was on Oct 20, and the repo hasn't changed since stars started climbing. In other words, almost nobody actually downloaded, ran, or raised issues with it.

Almost nobody knows about it overseas. Searches on X, Reddit, and Hacker News turn up no discussion; secondary spread is limited to a few Chinese aggregator sites and one personal notes library.

Conclusion: this is a sample of "one media push plus bought forks / bought stars," not organic growth. Whether the stars were all bought can't be checked one by one (GitHub no longer provides per-star timestamps), but the fork evidence combined with the steady curve makes it highly likely (inferred). Who bought the traffic can't be determined from the outside.

Whether the direction itself is interesting

Yes. The hardest part of long-form story video is continuity: by the third segment the cat has changed color and the lead is wearing different clothes. InWorld doesn't try to make the video model remember this. It puts an agent outside the model to keep a "ledger": fixed identity (reference images, voice) is managed separately from state that changes (position, mood, what's in hand), and every segment gets reviewed after it's generated. The same approach works for interactive short dramas and AI comic series.

But the competition on this path is heavy:

Building the "world model" itself means training models, stacking compute, and chasing evaluation leaderboards. Your opponents are Ant, Tencent, and Google. An individual or small team can't win stars at this layer, let alone real users.

What to learn, and what not to

Worth learning: the cold-start packaging of an AI lab. A technical report PDF, a polished project page with a demo video, a bilingual README, Hugging Face weights, and a feature in a top media outlet. Put together, it looks like a legitimate lab. Its website and README show 15 commits credited to "Codex", so this packaging was fast to produce.

Worth learning: the "agent keeps the ledger" structure. This layer doesn't require training a model. You can build it on off-the-shelf video APIs.

Don't copy: buying stars. Cross-check watchers, download counts, and fork timestamps and the pattern gives it away. Bought numbers also buy you nothing: no one downloads, no one opens issues, no one reposts. The project went quiet in Oct 20.

If it were you

Bottom line: Don't build an "open-source world model" after this one, and don't treat its stars as evidence of demand. The only part worth thinking about is the part it did badly: getting ordinary people to actually play it.

Who it's for: Indie developers comfortable with agents and video APIs who want an open-source project that can earn stars; small teams making interactive short dramas or AI comics who want an open "story consistency" layer.

Who it's not for: Anyone who wants to train their own video model and climb evaluation leaderboards (your opponents are big tech); anyone expecting 1,600 real users waiting for it (they don't exist).

How to do it: Build only the Murea-Mind layer: an open-source "interactive story director" agent with a world-state ledger, reference images to lock characters, rollback branches, and multiplayer rooms. Leave the visuals to existing video APIs (Veo, Kling, hosted LTX), so it runs without a single GPU. Put a dozen-plus turns of continuous story that plays directly at the top of the README, plus a live demo link. InWorld lacks exactly these two things: something that runs, and something people can play.

How to get users: Package it as a skill or MCP that Claude Code / Codex can install directly, and get it into awesome lists and skill directories. On English X, post a side-by-side clip of "the same cat staying unchanged across ten story turns": consistency is the easiest and most visual selling point in this category. In the Chinese community, pitch newsletters for features.

When it's too early: Interactive video has been a direction every major player has been shipping for 2026 years. Once a big vendor builds "world state" into its API, this layer gets eaten. Shipping within three months is fairly safe (my estimate).

Signs you've misjudged it: A month after launch, no one has played the live demo for more than five turns, or the stars come mostly from channels you pushed yourself, with no organic issues or meaningful forks. That means people don't actually want "watch AI film my story" that badly. Stop.