Emma Should Have Contributed to Italian Technological Sovereignty, but Can't Even Multiplicate by Nine
Emma Should Have Contributed to Italian Technological Sovereignty, but Can't Even Multiplicate by Nine
A simple elementary school question is enough to dismantle weeks of statements on Italian technological sovereignty. There are an awful lot of questions that have put Emma-5, the flagship language model of the Emma family developed by the company Egomnia S.p.A. and founded by Matteo Achilli, in difficulty.
The spark was a satirical post that went viral, where a user joked that Emma-5 had given them excellent financial advice, humorously presenting it as the artificial intelligence that will free Italy from the dominance of OpenAI and Anthropic. In reality, Emma-5 responded in a nonsensical and unintentionally ironic way to almost every question asked by users who, for personal amusement, wanted to test it out, at least until the tool was taken offline. As of writing, it is indeed no longer accessible.
The Technical Specifications Behind the Hoax
The official project page lists four models from the Emma family, with publicly verifiable numbers. Emma-3 has 125.3 million parameters and weighs 0.6 gigabytes: a toy model, useful only for rapid prototypes. Emma-4 increases to 317.1 million parameters and 1.48 gigabytes, designed for lightweight chatbots. Emma-5, the accessible model considered flagship, has 550.4 million parameters and weighs 2.46 gigabytes. Emma-6, still in training and not available to the public, is expected to reach around 1.55 billion parameters.
Architecturally, Emma-5 is a transformer decoder-only in the GPT style, with Grouped-Query Attention in a 4:1 ratio, RoPE for positional encoding, RMSNorm for normalization, and SwiGLU as the activation function. The vocabulary is based on a 50,000 token SentencePiece BPE tokenizer, and inference runs through ONNX Runtime, executable locally without connection to external servers. However, the context window stops at 2,048 tokens: a low threshold that significantly limits the model's ability to handle long conversations or extensive documents.
On the internal benchmark Vespucci, a list of 875 questions in Italian developed by the company itself, Emma-5 reports a score of 57.8 percent. This data comes from a proprietary test and not from an independent third party, so it deserves the same caution that is applied to any self-produced benchmark.
What Egomnia Says, and What It Doesn't Promise
One point in favor of the company is that the official website does not promise results that the product cannot deliver. The models of the Emma family are described as experimental projects, explicitly not designed to compete with high-end international models. The disclaimer published on the portal clearly excludes any professional, legal, medical, or financial use of the generated responses. The viral screenshots about "financial advice" hit precisely the only area that the producer expressly prohibits: much like mocking a pocket calculator for not driving a car.
How Much Does Emma Cost and How to Use It
Basic access to the web chat, available without registration, is free but with a message limit for each conversation. For those who want to remove that limit and get priority on the larger models, the Premium subscription costs 1 euro per year. Those who prefer to run the models on their own computers can download the weights for free from Hugging Face under the OpenRAIL-M license, or request API keys for business use. An app for iOS and Android, according to the website, is still on the way.
On the data front, Egomnia also sells EmmaSFT 6.0, a set of over 100,000 question-answer pairs used for fine-tuning the future Emma-6, distributed through a dedicated online store and made free for non-commercial universities and research centers.
The Founder with a Past Already Under the Spotlight
The name chosen for the project is not casual: Emma is Achilli's daughter, who founded Egomnia in 2012 while still a student of economics at Bocconi University in Milan. At that time, Egomnia was born as a professional social network, a sort of Italian alternative to LinkedIn based on a candidate ranking algorithm. The weekly Panorama Economy dubbed him the "Italian Zuckerberg" as early as 2012, followed two years later by the BBC docuseries The Next Billionaires and then a book published by Rizzoli.
In 2017, his story became a film, The Startup, directed by Alessandro D'Alatri, with Andrea Arcangeli portraying the protagonist. However, the theatrical release triggered a flurry of articles questioning the correspondence between the business collaborations touted by Egomnia and the real numbers of the company. Achilli rejected the challenges and attributed them to excessive media attention, before continuing to lead the company, now listed on Euronext Growth Milan with a capitalization that stops at just over 2.3 million euros.
This background weighs on public judgment more than the technical quality of the product itself: the widespread sensation online is that of a script already seen, and of ambitious proclamations followed by a product that fails to live up to the expectations created.
Those wanting to seriously discuss the independence of language models have, in Italy, two other references on a different scale. The first is Minerva, the family of models developed by Sapienza University of Rome, trained from scratch on Italian texts. The second is Modello Italia, the project of the company iGenius, built on computational resources an order of magnitude greater than those of Emma. Also, the European debate on sovereign language models, from the Swiss Apertus project to community initiatives, moves on parameters, data, and infrastructures that, to date, the Emma family does not possess.
Between Irony and Concern: Industry Reactions
Not all technical comments have limited themselves to mockery. On StartupItalia, a lawyer experienced in digital law and a lecturer at several Italian universities defined the affair as more worrying than comical. According to this interpretation, the risk is that a public company leads the public to confuse an educational scale experiment with the actual level achieved by Italian artificial intelligence, thereby weakening in the eyes of millions an idea, that of technological sovereignty, which remains commendable in itself.
In conclusion, the Emma case does not only concern a 2.46 gigabyte chatbot capable of running on any laptop: it also concerns how a technological project is communicated to the public, and how the disproportion between words and technical specifications can turn an initiative, perhaps honest in its declared limits, into the punchline of the day.