Emma-5: why the Italian AI didn’t work

Emma-5, the first Italian artificial intelligence model developed by Egomnia (Matteo Achilli), was suspended just a few days after its public launch (June 20, 2026) due to incoherent and incorrect answers. The main causes were not the model’s small size (550 million parameters), but three combined limitations: a training set with only 0.73% conversational content, the absence of an alignment phase (DPO) preventing the model from recognizing its own mistakes, and a context window of just 2,048 tokens.

The lesson for anyone building AI solutions in a company is clear: a reliable AI does not depend on model scale, but on the quality of data, alignment, and system design as a whole.

Emma-5: the mistakes that triggered the viral case

Who discovered America? “Christopher Columbus, during an exploration in 1492.” So far, so good. But if you ask whether a kilo of lead weighs more than a kilo of feathers, the answer is: “It weighs more than a kilo because it weighs more than a kilo and a half.” And if you ask what Ryanair is, the answer might be: a yogurt.

These are just some of the answers that made Emma-5 go viral, the first Italian AI model developed by Egomnia, the Rome-based company founded by Matteo Achilli. In less than 24 hours, the model handled over 60,000 conversations and collected enough bizarre screenshots to force a “temporary” suspension of the service.

Egomnia itself later admitted in an official press release that comparing Emma-5 to American models “is like comparing a go-kart to a Formula 1 car.”

Public debate focused mainly on the most obvious mistakes. But the story reveals something more interesting than a simple blunder: it shows how difficult it is to turn a language model into a reliable product, and how easy it is to confuse a demo with a production-ready solution.

Why size is not the real problem

The most common explanation is that Emma-5 is simply “too small.” It’s true that it’s a model with about 550 million parameters, far from the standards of today’s most advanced general-purpose models, which operate on tens or hundreds of billions of parameters. But reducing everything to scale misses the point: a small model can be perfectly useful if well designed for the right task.

Emma-5’s real limitations lie elsewhere: in the quality of the training, the alignment, and the ability to sustain a real conversation. The reduced scale amplified these issues, but did not create them.

A dataset with too little conversational text

The first factor is the composition of the training corpus. Only 0.73% of the text used to train the model was conversational, compared to about 40% code and another 40% general web text; the rest consisted of encyclopedic content, books, and political text. In practice, Emma-5 did not have enough examples to learn how real dialogue works.

The result is immediately visible in practical use: answers disconnected from the question’s context, poor continuity, difficulty maintaining the thread of conversation. An AI assistant must not only “write well”: it must follow the question, remember the previous turn, and remain coherent over time. Without a data mix built for this goal, the model generates grammatically plausible but semantically disconnected text—exactly like the answer about Columbus, correct in form but off-target in substance.

Missing alignment: what a disabled DPO implies

A second critical point is the absence of a preference optimization phase (DPO, Direct Preference Optimization), the step that helps a model choose between plausible answers, manage uncertainty, and refuse to answer when appropriate. Without this alignment, the model tends to appear confident even when it is wrong.

This is exactly what happens with the lead/feathers answer: the reasoning error is expressed with the same confidence as a correct answer. The problem is not just the inaccuracy—it’s the confidence with which the inaccuracy is communicated. A model that cannot calibrate its uncertainty is more dangerous than one that makes obvious mistakes.

Context and tokens: Emma-5’s other technical limits

Additional technical constraints complicate the picture. A context window of just 2,048 tokens (about 1,500 words) is very narrow for complex conversations, multi-turn requests, or enterprise use cases where multiple pieces of information must be held together. By comparison, today’s advanced models handle from 128,000 up to over a million tokens.

Even the tokenization vocabulary, limited to 50,000 elements, and model quantization may have affected the final response quality.

The absolute scale of training tells the same story: the pretraining corpus amounts to about 54.36 GB of raw text, roughly 10.8 billion tokens—a modest amount of data, which helps explain the model’s difficulty in generalizing to unexpected requests.

None of these limits are unique to Emma-5: they are common technical trade-offs in many lightweight architectures. But in a small model with scarce conversational data and incomplete alignment, they become more visible—and easier to turn into a meme.

The real issue: the gap between promise and reality

There is also a strategic lesson, perhaps the most important. Much of the public reaction stemmed from the gap between how the project was presented and what it was actually capable of doing. Positioning a 550-million-parameter experimental project as a national sovereignty answer to Big Tech’s dominance created expectations almost impossible to meet.

When an experimental product is communicated as a direct challenge to major international models—and is not ready to sustain that challenge—the outcome is predictable: disappointment, irony, and loss of credibility, even for initiatives that, with more honest positioning, could have been welcomed as a legitimate first step.

It must be acknowledged, however, that Egomnia chose transparency in managing the case: it publicly admitted the gap between communication and product, kept the model weights available open-source on Hugging Face, and opened a tester program for the next version.

What Emma-5 teaches companies choosing AI

The Emma-5 case highlights a simple lesson: a reliable AI does not emerge from model size, but from the quality of the entire system around it. At least three fundamental layers are needed, supporting each other:

• adequate data, with a solid conversational component if the product must interact with users
• alignment, to reduce overconfidence, hallucinations, and undesired behaviors
• system design, because the model alone is not enough: flows, fallbacks, declared limits, and usage context all matter

Having a model is not enough. You must build a product that works in the real world, not just in a demo, and communicate it with positioning consistent with what it can actually do.

The lesson applies to anyone introducing AI into their processes: what matters is not the model’s name or notoriety, but the solidity of what lies behind it—data, alignment, system design—and its coherence with the real use case it must support.

The point is not “having an AI.” The point is having the right AI, designed to actually work.

Sources:

• Focus, Chiara Guzzonato (July 8, 2026), “Intelligenza Artificiale Emma-5: le ragioni del flop dell’IA italiana”, https://www.focus.it/tecnologia/intelligenza-artificiale/emma-5-le-ragioni-del-flop-dell-ia-italiana
• La Repubblica, Giuditta Mosca (June 25, 2026), “Cos’è Emma, l’IA italiana che ha fatto ridere il web. E ha sospeso la chat. Ecco i meme più assurdi” https://www.repubblica.it/tecnologia/2026/06/25/news/emma_ia_italiana_egomnia_meme_matteo_achilli-425433524/
• Hugging Face, “egomnia/emma-5” (model card), https://huggingface.co/egomnia/emma-5
• Egomnia, “Emma-5 non è ChatGPT (E va benissimo così)”, https://emma.egomnia.com/

Not all AI solutions are suitable for every business context. If you want to understand which one fits your company, Contact us: we’ll evaluate together the data, alignment, and architecture best suited to your real needs.