AI Radar checks GitHub, Hugging Face, arXiv and Hacker News once a day and keeps a snapshot of each project's stars, likes, upvotes and points. A repo, its Hugging Face models and spaces, and the threads about it are merged into one project, so you see every signal in one place.
Top picks and Just found
Just found projects were discovered in the last few days and are still being checked. Once they show steady growth and pass spam, fork and quality checks they become Top picks , the default view. Projects that go quiet move to Archived , which search still covers.
Momentum vs growth
Momentum (0 to 100) is the ranking number. It combines how fast a project is growing on every source it appears on, with a boost for being new and for being discussed in more than one place. It is the default sort, and the boxed number on each breakout row.
The growth tier (Calm growth , Rising, Hot, Viral) measures only one thing: how many stars, likes or points a project gains per day. So a project launched yesterday, or one with a busy Hacker News thread, can have a high momentum while its growth is still calm. Where we know why, the momentum tooltip and the details say so, for example "Boosted because it's brand new" or "Strong momentum from launch buzz; growth is steady", and cards add a short new or buzz next to the number.
Breakouts this week
The highest-momentum Top picks launched in the last 7 days that are Rising or faster , with a one-line description of each. Calm and too-early projects are left out, so on a quiet week (or with narrow filters) there may be fewer than six, or none. Breakouts follow your filters, including the launch window.
Launched: last 90 days by default
AI Radar is about what's new, so the Projects list starts with projects launched in the last 90 days . The filter shows as a pill above the results: remove it (or pick "Any time" under Launched) to include older projects that are still growing. Counts, Breakouts and the tab number follow the same window.
Growth per day
How many stars, likes, upvotes or points a project gains per day right now , shown with its unit (for example +204 stars/day ). A repo with 300 stars that gained 120 yesterday is more interesting here than one with 30,000 stars that gained 40.
Acceleration (×)
A green 2.4× means the project is growing 2.4 times faster than its earlier pace: the last day's gain compared with the day before. It appears from 1.15× up, is shown as 10×+ above ten (a jump from a tiny base says little beyond "much faster"), and is hidden when a project gains only a few stars, likes or points a day.
Measured vs. estimated
Once we have seen a project in at least two snapshots, growth is measured from the real difference between them. Before that it is estimated as the lifetime average (total divided by age), and marked est.
Too early to tell
Projects we haven't measured over enough snapshots show Too early to tell instead of a tier or a growth number. They never appear in Breakouts or under a tier filter until real growth has been measured.
Tiers
Calm growth Normal, steady pace: most projects. They are listed, but never shown in Breakouts.Rising Faster than most new projects.Hot Well above the usual pace.Viral Among the very fastest right now.
The filters Rising & up and Hot & up include every faster tier too. Tiers measure growth only; see Momentum vs growth .
Sources and their numbers
GitHub : stars (and forks in the details)Hugging Face (HF): likes (and downloads in the details) on models, spaces and datasetsPaper (arXiv): the linked paper. Its number is upvotes on Hugging Face Daily Papers, shown as HF Papers 29 upvotes (arXiv itself has no votes)Hacker News (HN): points on the discussion thread
Cards spell this out, for example HF 111 likes . A source can be linked before we have a number for it; the details then say "linked, no numbers yet".
Quality score
Quality (0 to 100) is how trustworthy a project looks: the author's track record, a licence, a real README, genuine activity, and spam, fork and copy checks. It decides whether a project can become a Top pick, and it is separate from momentum: a quiet, solid project can score high on quality and low on momentum.
Some projects are also checked by an AI judge : a second model re-reads the project and reports how confident it is in the main category and how likely machine learning is essential to it. The details show this as one line in the quality section; it is a sanity check, not a human review.
Early gems
Launched in the last 3 weeks, still below a "small" total for its source, and already Rising or faster. These are the ones most people haven't found yet.
Variants and flags
Quantized, fine-tuned and merged copies of a model, adapters and re-uploads in another format are grouped under the original and shown as +N variants . In the details, Sibling models (other sizes or versions from the same publisher) are listed apart from these community variants. Muted labels such as tutorial or wrapper are informational: they describe a project, they don't hide it.
AI-powered apps
Most projects here are AI building blocks: models, agents, tools and infrastructure. Some are apps whose key feature relies on machine learning, such as a photo editor built around a segmentation model. Those carry a muted AI-powered app label. It is informational and does not change the ranking.
Categories, including Skills & Plugins
Every project gets one main category; Skills & Plugins also lists domain-specific skills. It covers agent skills, MCP servers and plugins for IDEs and AI assistants: small add-ons that give an assistant a new ability, as opposed to full agent frameworks (Agents & Automation ). A skill for a particular domain, such as a Blender MCP server, keeps that domain as its main category (3D) and is listed under Skills & Plugins as well, so the Skills count is larger than the number of projects whose main category is Skills. The category's tooltip says how many such domain skills it includes.
Research count
The Research tab includes Hugging Face Daily Papers as well as the arXiv crawl, so it lists more papers than the arXiv source count. Each paper says where it was found.