Italy's Domyn releases 4-billion-parameter small model "Domyn Edge," disclosing training checkpoints and data

Support in inference software came first, official release the next day

Italian AI company Domyn released "Domyn Edge," a small model with 4 billion parameters, on Oct. 20. On the previous day, Oct. 19 (Oct. 20 Japan time), @aitrackerbot, an X account that tracks the appearance of models, had reported that support for Domyn Edge was added in advance to "vLLM," inference software for running AI models.

Domyn is the renamed Italian AI company iGenius. In June, the European Commission selected "EUROPA," a consortium led by Domyn, as a winner of a call to develop large European AI models. EUROPA aims to build an open-source model that supports all 24 official languages of the EU. Backers of the company include U.S. financial giant BNY and Abu Dhabi's G42.

Disclosing intermediate models, data and procedures

Domyn Edge was trained from scratch on 13 trillion tokens of data. While many companies release only the weights of a finished model, Domyn will also release models from intermediate stages of training (every 1 trillion tokens), training data, and training procedures and tools. This release is the first installment, and the company said it will publish a model that has gone through the final stage of training, along with a technical report, in the coming weeks. The aim is for European companies and research institutions to use it as a foundation for building their own models.

Training used "Leonardo," a European supercomputer operated by Italy's computing center CINECA. Pre-training took about 2.3 million GPU hours, and the whole process about 5 million GPU hours. With 4 billion parameters, the model can run on a single GPU. It was trained on contexts of up to 1 million tokens, and the company recommends up to 256,000 tokens for practical use.

Beats most open models of the same size, but trails Qwen on some tests

In Domyn's evaluations, the version before final training beat U.S.-based Hugging Face's "SmolLM3-3B" on 27 of 28 tests. Results against China's Alibaba's "Qwen3.5 4B," however, were mixed. On MBPP, a code-writing evaluation, Domyn Edge scored 96.6, above Qwen's 92.8, but on IFEval, which measures instruction following, it scored 90.2, short of Qwen's 92.6. On an evaluation of finding information in long contexts (RULER, 128,000 tokens), it scored 78.0, far below Qwen's 93.2.

Evaluations self-measured, license not named

All figures were measured by Domyn itself, and there are no external evaluations yet. On terms of use, the company describes an "open license," but its official article does not name the specific license.