Domyn Opens Up Edge, a 4B Agentic Model

Researchers usually see only the finished version of an AI model. With Edge, Milan-based Domyn is publishing the steps along the way. The company has begun what it calls “a staged open release” of the 4-billion-parameter agentic model, sharing “its intermediate training checkpoints, recipes, and tools” alongside it, according to its announcement on 20 October 2026.

Edge was “trained from scratch on CINECA’s Leonardo supercomputer”, the Italian public system. Domyn presents the release as a way to stretch the value of European public compute: rather than handing over one finished model, it wants others to pick up training part-way and build their own specialist versions.

Checkpoints arrive first, post-training into November

The rollout starts with checkpoints from pre-training, mid-training and supervised fine-tuning, and “will progress along different post-training stages into November 2026,” Domyn says. The first batch is on Hugging Face, with the training code released as open source on GitHub.

Domyn gives a banking example: a bank could take an earlier checkpoint and adapt Edge to its own terminology and transaction-monitoring processes, either with its own AI team or with Domyn. The company says details of checkpoints, recipes, supported hardware and the final release roadmap will follow on its site.

Domyn calls Edge “the best-performing fully open model of its size”, saying published benchmarks place it ahead of open models in its category. The announcement links to a paper but quotes no benchmark scores and names no rival models, so the claim is Domyn’s alone for now.

A small model meant to run on site

Domyn says Edge is designed for multi-step agentic work and tool use. At 4 billion parameters it is cheaper to adapt and run than larger models, and it can run on compatible local hardware. The company points to settings such as manufacturing, where substantial GPU infrastructure would be impractical, and says Edge can also work alongside larger models within an agentic system.

“Sovereignty is about control,” said Luca Antiga, Domyn’s chief technology officer. “Europe needs an ecosystem of builders, and the real potential of Edge is in what the ecosystem will be able to create, together with us or independently.”

Part of a bigger sovereign push

Domyn builds AI for regulated industries and lists the Italian government, Fincantieri and BNY, also an investor, among the organisations it works with. Eurizon Capital, Rabobank and Abu Dhabi’s G42 are among its other backers. It is also developing Colosseum, which it describes as Europe’s largest AI supercomputer, with NVIDIA and G42. In June, the European Commission picked a Domyn-led consortium, EUROPA, to build an open-source model covering all 24 official EU languages, the company says.

Publishing checkpoints and recipes is rarer than releasing final weights, and more useful to anyone who wants to train a specialist model without starting from zero. Domyn is showing Edge at NVIDIA GTC in Berlin. The open question is whether the post-training releases due by November, and the benchmark detail in its paper, back up the “best-performing” label.