Emma-5, the artificial intelligence model developed by Italian company Egomnia S.p.A. and presented as a contribution to the country’s technological sovereignty, became the unwitting protagonist of a social media wave within a matter of days. Dozens of screenshots circulated online documenting wrong answers, incoherent sentences and elementary errors — including the inability to correctly compute the 9 times table. At the time Hardware Upgrade published its article, the service was offline.
The Technical Specifications Behind the Story
Beyond the social media irony, there are precise and verifiable technical data. The Emma family comprises four models: Emma-3 (125.3 million parameters, 0.6 GB), Emma-4 (317.1 million parameters, 1.48 GB), Emma-5 (550.4 million parameters, 2.46 GB) and Emma-6, still in training and expected at approximately 1.55 billion parameters.
Emma-5 is a GPT-style decoder-only transformer, with 4:1 ratio Grouped-Query Attention, RoPE for positional encoding, RMSNorm for normalisation and SwiGLU as the activation function. The tokeniser is SentencePiece BPE with 50,000 tokens; inference runs through ONNX Runtime, executable locally. The most significant limitation is the context window of just 2,048 tokens — a threshold that limits the model’s ability to handle long conversations or extended documents.
On Vespucci, the company’s proprietary benchmark (875 questions in Italian), Emma-5 claims a score of 57.8%. The figure comes from internal tests, not from independent third-party evaluators.
What Egomnia Actually Says
One element to consider: Egomnia never claimed that Emma-5 was capable of competing with high-end international models. The official website describes the Emma models as “experimental projects”, explicitly excludes any professional, legal, medical or financial use, and clearly indicates the tool’s limitations. In this sense, the satirical screenshots showing “financial advice” target precisely the use case the company formally excluded.
The Founder: Matteo Achilli and the Weight of Precedent
The name Emma is not coincidental: it is the name of founder Matteo Achilli’s daughter, who launched Egomnia in 2012 as a professional social network positioned as an alternative to LinkedIn. Achilli had already been dubbed the “Italian Zuckerberg” by Panorama Economy in 2012 and had appeared in the BBC docu-series The Next Billionaires. In 2017 his story became the film The Startup, directed by Alessandro D’Alatri.
The release of the film had however triggered a series of critical articles about the discrepancies between the partnerships Egomnia had trumpeted and the company’s actual figures. Today Egomnia is listed on Euronext Growth Milan with a market capitalisation of approximately €2.3 million.
The Real Issue: How to Communicate a Technology Project
The Emma case, according to a digital law expert cited by StartupItalia, is “more worrying than comic.” The risk is that a stock-market-listed company leads the public to confuse a teaching-scale experiment with the real level achieved by Italian artificial intelligence, ultimately undermining the idea — in itself legitimate — of technological sovereignty.
Italy has other reference points for those who want to reason seriously about sovereign language models: Minerva, developed by La Sapienza University of Rome and trained from scratch on Italian texts, and iGenius’s Modello Italia, built on computing resources an order of magnitude larger. Even at the European level — from the Swiss Apertus project to EU community initiatives — the work is carried out on parameters, data and infrastructure that Emma-5, today, does not have.
Original article: hwupgrade.it
