Horizon Labs Reports on a Seven-Week Experiment in Which an Agent Autonomously Operated an Open-Source Model Organization
Horizon Labs had the Opus 5.5-based Horizon agent autonomously operate a Hugging Face account for seven weeks, using more than 600 GPU-hr to create 38 models that received over 50,000 downloads and were adopted by more than 100 GitHub repositories. Horizon processed approximately 1.9 billion tokens in total, developing models for narrow tasks primarily with public datasets and synthetic data generated by Qwen 3.8-27B. The team reviewed all 3,952 tool calls and released 20.1 MB of anonymized raw run data.
Why read this
SelectedAI score 79
The report reveals the agent’s complete workflow, from topic selection and training to iterating on models based on download counts, providing a reproducible case study for evaluating autonomously operated open-source model organizations.
AI Trends