Analysis · 8 min · August 20, 2026

3.5 million downloads, almost no footprint: the Ornith-1.0-35B question

The repository ornith-ai/Ornith-1.0-35B-GGUF went up on July 18 — the same day Kimi K3's weights dominated every AI headline, which is probably why nobody noticed. In the latest trending snapshot on this site it had accumulated 3,544,218 downloads and 1,053 likes. That download figure puts it ahead of every model in the top sixteen except four repositories that have been accumulating traffic for longer: gpt2, gpt-oss-20b, Llama 3.1 8B and DeepSeek R1. A brand-new 35B model from an unfamiliar account outdownloading GLM-5.2 and DeepSeek V4 Pro in five weeks is either a genuinely remarkable launch or a number worth a second look. We went looking.

The numbers that don't add up

Downloads and likes on the Hub come from overlapping but different populations: likes require someone to be browsing the repository page and choosing to click something, while downloads can come from a single human, a CI pipeline, a mirror, or a benchmark harness pulling the same weights on a schedule. For organically popular repositories the two numbers still tend to move together, because the same wave of engaged users drives both. Set the ratios side by side and Ornith looks out of place. DeepSeek-V4-Flash-0731 carries about one like per 850 downloads. GLM-5.2 is closer to one per 530. Ornith-1.0-35B sits at roughly one like per 3,370 downloads — a ratio several times thinner than any comparable entry in the same snapshot.

None of that proves anything by itself. Some genuinely useful low-level or infrastructure repositories get pulled by automated pipelines far more than they get liked, and a thin like ratio is not evidence of wrongdoing. It is, however, exactly the kind of anomaly this site's editorial principles say we should attribute rather than wave past — so here it is, attributed: a real number, from our own snapshot, that we cannot fully explain.

What the repository does and doesn't tell you

We read what's actually there. No written model card describing training data, architecture family, or tokenizer. No benchmark numbers, vendor-reported or otherwise. No paper, no organization page beyond the single repository, no discussion thread with more than a handful of comments. The GGUF quants were uploaded directly by the account rather than following the usual path of a base checkpoint that a known community quantizer — bartowski, unsloth, mradermacher — converts after the fact. That's not inherently suspicious; plenty of labs ship their own GGUF builds now. It is one more data point in a repository that otherwise offers almost none.

One thing does check out: the license file in the repository metadata does say MIT, which at least means the permissive claim isn't fabricated. But “the license is real” and “the model is what it appears to be” are different claims, and we can only verify the first one from where we sit.

How this one reached us

We didn't find Ornith by browsing. This site runs a weekly catalog-freshness workflow that flags any trending model whose author matches none of the families in our own coverage list — ornith-ai isn't an alias of anything we already track, so it cleared that bar cleanly, and its like count comfortably cleared the threshold the workflow uses to separate a real candidate from noise. Monday's run filed it as a new-family candidate, which is the entire reason this post exists. The automation did its job; this is us doing ours.

What we're doing about it

Nothing, yet — and that is the point. This site's rule for the directory and the picker is that an entry gets added after we've checked the license and put real hardware numbers behind it, not from a download count alone. We ran a handful of prompts against the GGUF ourselves; a handful of prompts is not a benchmark suite, and we're not going to publish a quality verdict built on it. Until Ornith clears the same bar every other entry in the directory clears — a real model card, or enough independent testing that its absence stops mattering — it stays out.

The broader takeaway travels beyond this one repository. A trending position on the Hub is one weak signal among several, not a verdict. Before you trust a sudden appearance at the top of any list — this site's trending page included — check whether the likes track the downloads, whether the account has a history beyond a single upload, and whether anyone has written down what the model actually is. That's worth doing for any model. It matters more for anything you'd run with file access or tool permissions. We'll update this post if ornith-ai publishes a model card or if independent testing turns up something concrete. Until then, the download count buys curiosity, not a recommendation.