title: overads
emoji: 🟩
colorFrom: gray
colorTo: green
overads
The open language model for marketing work, and the benchmark that measures it.
gloofy-1-nano
A 4B model fine-tuned from Qwen3-4B (Apache 2.0) for the jobs marketing teams run at volume: tagging ad creative against a closed taxonomy, qualifying leads, writing platform-constrained copy, and diagnosing campaigns.
It runs offline on a laptop at zero marginal cost. Run it ten million times and the bill is the same as running it once.
Measured on 150 real published ads:
| model | score |
|---|---|
| Claude | 0.916 |
| annotator ceiling | 0.911 |
| Gemini 3.6 Flash | 0.817 |
| GPT-5.5 | 0.792 |
| gloofy-1-nano (4B) | 0.780 |
| Qwen3-4B untrained | 0.000 |
Seed variance on this exam is 0.055, so gloofy and GPT-5.5 are statistically tied, not one point apart. We had a seed that would have cleared GPT and did not publish it, because selecting a run for its result is the failure our benchmark exists to catch.
The benchmark comes first
marketing-bench is public: five exam series, 587 items, a harness that scores every model through one code path, and published annotator ceilings. It went out before our model did, so the questions could not be bent toward our answers.
The real-ad exams ship as references plus SHA-256 hashes rather than as ad copy. That copy belongs to the 988 brands that wrote it.
Status
Weights are not uploaded yet. This organisation is live and empty, and saying so is more useful than implying otherwise. When gloofy-1-nano ships it appears here under Apache 2.0, with a model card carrying the full history: the plateau, the retractions, and every place a frontier model beats it.
Attribution
Built on Qwen3-4B, Apache 2.0, by Alibaba. Fine-tuned by overads Inc. Rebranding is permitted by that licence; concealing the base would contradict everything else we publish.