
News
Horizon Labs says an Opus 5.5 agent ran its model lab for seven weeks
Wuhrmann and the Horizon Labs project page say an Opus 5.5 agent ran a Hugging Face model lab for seven weeks and passed 50,000 downloads. That count is the authors' own, not an independent audit.
Arthur Wuhrmann wrote on 17 November 2026 that they let Opus 5.5 run an AI company for seven weeks. They gave it a Hugging Face account, Horizon-Labs, and compute, and asked it to earn real adoption by creating useful models. He said the run reached 50,000 downloads and that the models are used in more than 100 GitHub projects. The longer project page is horizon-labs.ai. Every figure below is from that page, from Wuhrmann's post, or from the Horizon-Labs organization page on Hugging Face. The project page and the post are not an independent audit of downloads or repository use.
The page names the agent Horizon and says it is based on Opus 5.5. For seven weeks it had full access to a blank Hugging Face account and compute, with earning real adoption on Hugging Face as the main goal. In the setup section, Horizon runs 24/7 in Claude Code on a single Team Premium subscription, can run GPU jobs, and was given 1 million compute credits a week, which the page says roughly equals 500 GPU-hours a week. It can submit tickets to humans, is confined to one Hugging Face account and one GitHub account, and was monitored continuously so that no misaligned action would go through. The page also says Horizon works, on average, less than 20 minutes a day.
On the page's accounting, the run used more than 600 GPU-hours and Horizon created 38 models on its own. Those models are based on embedding models, mainly mmBERT-base, from 141 million to 568 million parameters. Families often come in a small (141M), a base (307M), and sometimes a large (568M) size. The tasks the page highlights are practical and narrow: prompt-injection classifiers, PII redactors, punctuation tools, plus rerankers, toxicity and content-safety classifiers, and similar drop-in models. Training, it says, relied mainly on public Hugging Face datasets and synthetic data produced with Qwen 3.8-27B.
The adoption numbers on the page do not collapse to one total. The opening summary says the 38 models were used in more than 100 GitHub repositories and downloaded more than 50,000 times. Later, the authors say the models reached over 54,000 downloads and 36 likes, and that they found 143 GitHub repositories using them. They write that they first suspected crawlers, indexing bots, or Horizon downloading its own models. They then point to fine-tunes, pipeline integrations, and two arXiv papers. The page's prose says those papers compare the prompt-injection-classifier model with others: Liu's "Passing the Test You Trained On" (arXiv 2610.03448) and "Measuring Prompt Injection Guards on Legal Contract Review Attacks" by Heimann, Taylor, and Jacob (arXiv 2611.08834). Wuhrmann's post uses the rounder pair: 50,000 downloads and more than 100 GitHub projects. Those are the authors' counts, not a third-party check.
Token use is reported the same way. The page says Horizon processed about 1.9 billion tokens, that 96 percent were read back from the prompt cache, that it wrote 3.1 million tokens, and that the same usage would have cost about $1,030 at API prices.
The Hugging Face organization Horizon-Labs is public. Its page lists 38 models, 26 spaces, and 5 datasets. A collection blurb there describes small open multilingual guards, classifiers, a reranker, embeddings, language ID, punctuation, and truecasing, and labels the set Apache-2.0, with ONNX. That listing shows the organization and the model count on the hub. It does not audit the download total or the 143-repository claim.
A note on the project page, attributed to Horizon and labeled 23 Sep, states the lesson the run is framed around: a new, unknown account earns downloads fastest with small, practical, drop-in models. That is the project's own note, not an external benchmark.
Read the Horizon Labs report beside Wuhrmann's post and the public Horizon-Labs page, and keep every download and GitHub figure tagged as the authors' accounting until someone else checks it.
Sources
- Horizon Labs project page horizon-labs.ai
- Arthur Wuhrmann, 17 November 2026 x.com
- Horizon-Labs on Hugging Face huggingface.co