Realtime AI News
An AI agent published 38 small models in seven weeks, with about 55,000 downloads
A research effort spanning EPFL, the ELLIS Tübingen Institute, the Max Planck Institute for Intelligent Systems and Imperial College London published 38 models built by an AI agent. The agent, Horizon, ran on Claude Opus 5.5 with only an empty Hugging Face account and roughly 500 GPU hours a week, producing small guardrail models for about $1,030 in token costs.
A research effort spanning EPFL, the ELLIS Tübingen Institute, the Max Planck Institute for Intelligent Systems and Imperial College London has published 38 models built almost entirely by an AI agent, according to a November 19 report from 36Kr's New Zhiyuan.
The agent, Horizon, runs on Claude Opus 5.5 and was embedded in Claude Code to operate around the clock. It started with nothing but an empty Hugging Face account and a compute quota of roughly 500 GPU hours a week, and the team gave it a single instruction: build models that developers would actually want to use.
Horizon behaves less like a researcher than a shrewd contractor. It wakes when Hugging Face refreshes download statistics, checks the leaderboard and only polishes models that show real demand, spending on average about 16 minutes a day on actual work. Before starting a project it weighs whether similar models are being downloaded, whether the data, base model and code can be reused, and whether developers can swap it in seamlessly.
All 38 models are fine-tunes of the multilingual base model mmBERT with parameter counts strictly between 141 million and 568 million, small enough to deploy cheaply on modest hardware. Data was generated without manual effort: a 27B open-source model produced synthetic batches, and Gemma 4 31B labeled and expanded the dataset.
The demand Horizon targeted was mostly guardrails for AI applications: prompt-injection detection, personal-data masking and harmful-content classification. The prompt-injection series alone drew about 17,000 downloads, with the basic version above 7,000, and one model adds punctuation to text in 88 languages, turning dense meeting transcripts into clean sentences.
The cost accounting is the striking part. Over seven weeks Horizon ran for just over 600 hours, using less than 20% of its quota, and consumed about 1.9 billion tokens, 96% of them prompt-cached, for roughly $1,030 at API prices excluding GPU costs. Human intervention was concentrated in the first three days; after October 22 the team stepped back entirely, handling only work orders Horizon could not resolve.
The team also published its agent's more human failures. On September 24 it accidentally made an account copy public, in mid-October it answered a work order incorrectly, and on October 22 a framework failure made it oversleep; it also downloaded its own model 42 times across six operations. Even after subtracting self-generated, internal-test and cross-promotion traffic, the researchers estimate more than 50,000 real downloads remain.
The results have spread: more than 100 GitHub projects integrated the models, two arXiv papers used them for evaluation, and the Hugging Face community built 18 derivatives. The lesson is that autonomous AI R&D does not have to mean burning capital on compute; it can start from developers' everyday needs. What to watch is whether Horizon can iterate on user feedback and feed those capabilities back into itself.
Why it matters
It offers an auditable, low-cost sample of AI building AI: 38 genuinely published models with downloads and integrations rather than a purely theoretical paper. If the download figures hold up under independent verification, small-model factories could become a routine supplier to the open-source community.
Nearby Updates
All11/19, 19:00
Alibaba says AI demand lifted cloud sales 53% as revenue rises 10%
Alibaba said growing enterprise adoption of its AI pushed cloud sales up 53%, while the South China Morning Post reported cloud services driving a 10% rise in overall revenue. Together the figures show AI demand converting into cloud revenue growth rather than remaining a research line item.
11/19, 17:50
ServiceNow report: 92% of Indian organisations hit by AI agent security incidents
A ServiceNow report says 92% of Indian organisations experienced an AI agent-related security incident in the past year, according to varindia. The figure puts agent security at the centre of enterprise deployment decisions, though the published summary leaves the survey’s sample size, industries and incident definitions undisclosed.
11/19, 15:01
Kling Image 4.0 ships native 8K generation and 2,000-character text rendering
Kuaishou's Kling has released Kling Image 4.0, an image model that generates native 8K output with detail that still holds up under heavy zoom, according to a QbitAI hands-on test. The update also adds up to 2,000 characters of text per image, more than 70 preset fonts, series of up to 10 linked images and long e-commerce detail pages.
11/19, 14:57
AURA 2.0 lets the model start the conversation in streaming AI
AURA 2.0 has been released with model-initiated conversations inside streaming AI, according to Blockchain News. The change moves part of the right to start an interaction from the user to the system, which reshapes both the product and its boundaries.