How trckr works
trckr.ai exists so you can stop refreshing lab blogs and model lists. It watches 64 public sources around the clock and tells you about new AI models within seconds.
1. Watching
A worker checks these sources on a schedule, using HTTP caching (ETag / Last-Modified) so unchanged pages cost almost nothing:
- Provider model API: How developers plug a model into their own apps and pay per use. A model that's "in the API" can be used by programmers, even if it isn't in a chat app yet. (OpenAI, Anthropic, Google, xAI, Mistral and others) every 30–60 seconds. A Model ID: The exact name developers type to use a model, like gpt-5-mini. New IDs often show up in a lab's systems before the official announcement, which is how trckr spots many releases early. often shows up here before the announcement post.
- Hugging Face: The website where labs and researchers publish models anyone can download. Think of it as an app store for AI models. organisations of the major labs, for Open weights: The model's files are public, so anyone can download it, run it on their own computer and build on it, usually for free. trckr labels these "Open". releases.
- OpenRouter: A service that gives developers one place to use hundreds of models from many labs. New models often appear there early, sometimes anonymously.'s public model list, which also reveals anonymous Stealth model: A model released anonymously under a code name, so its maker can test it with real users before the official launch..
- Official blogs, changelogs and model documentation pages.
- Model-support pull requests in popular Inference: Running a trained model to get an answer. Training builds the model once; inference happens every time someone uses it. libraries (Early signal: A hint that a model is coming before it's announced: a name spotted in a lab's systems, an anonymous test model, or code that adds support for it. Usually right, sometimes not.).
- Tech press and community feeds for the news brief.
2. Understanding
Every new item becomes a signal. Signals about the same model are merged into one release, so eight sizes and three Quantisation: Shrinking a model by storing its numbers less precisely, so it fits on smaller computers. A little less accurate, much lighter. FP8, GGUF, AWQ and MLX are common formats. copies of a new open model are one alert, not eleven. An AI classifier names the release, decides whether it's a language, image, video or audio model, whether it has Open weights: The model's files are public, so anyone can download it, run it on their own computer and build on it, usually for free. trckr labels these "Open"., and how important it is (major, notable or minor). Major is kept for new general-purpose models from the leading labs, the kind behind ChatGPT, Claude or Gemini. Anything not yet officially announced is an Early signal: A hint that a model is coming before it's announced: a name spotted in a lab's systems, an anonymous test model, or code that adds support for it. Usually right, sometimes not., listed apart from the releases until it's confirmed.
Before a model is called new, trckr checks whether it was already public: its files on Hugging Face (the first community post shows when they were out), OpenRouter's listing date, the lab's own changelog and the date in the model's name. A release is dated by the earliest of these. If trckr only finds a model a day or more after it came out, the model goes on the site at its real date and nobody gets an alert for old news.
It also writes one sentence in plain English for people who don't follow AI, and notes where you can try the model: in an app, for developers, or as a free download. The glossary explains the jargon.
3. Notifying
Each release goes to every channel whose filter matches: browser and phone push, Telegram, Discord, email or ntfy. Each channel has its own switches for model types, labs, importance and early signals, so you can get everything in Telegram and only the big launches by email.
The crawler
trckr's requests identify themselves with the user agent trckr.ai/0.1 (+https://trckrai.com/about). It only reads public pages and APIs, polls at modest intervals with caching, backs off on errors, and skips sites that block automated clients rather than working around them. When a story comes from Hacker News, it reads the start of the linked article once to find the date it was published, and only where the site's robots.txt allows it (rules for trckr, otherwise for *).
Mistakes happen
Early signals are sometimes wrong, and the classifier can misjudge a release. Corrections are made by hand when spotted. Treat alerts as a heads-up and check the linked sources.