Emma Is Back: Egomnia Opens the Emma-6 Beta With a Model Almost Three Times Larger

Emma is back: Egomnia opened a beta of Emma-6, the new version of its Italian large language model, in an email to registered testers on the evening of 12 October, with 1.55 billion parameters, training carried out entirely in Italian and a separate safety filter called Emma Guardrails 1.0. The filter was trained on more than 44,700 real prompts that Egomnia collected during the launch and use of Emma-5, the version it suspended in June. The Bright Minded received the announcement by email as a registered beta tester, and the parameter count and training details do not yet appear on Egomnia’s public pages.

What happened to Emma-5

Egomnia, the company founded by Matteo Achilli and listed on Euronext Growth Milan, suspended Emma-5 on 25 June, five days after launch, once public testing had exposed errors in basic reasoning and facts. The model ran on roughly 550 million parameters with a context window of 2,048 tokens, the amount of text it could consider at once, and its launch language had implied far more. The Bright Minded argued at the time that the instinct to build Italian artificial intelligence (AI) capability was sound and that the execution earned the embarrassment.

What Emma-6 changes

Emma-6 is roughly 2.8 times the size of its predecessor. Egomnia’s email describes it as still a relatively small model, while claiming a clear improvement in how well it understands requests and in the quality of its answers. The company says it trained Emma-6 only on Italian, in line with its aim of a model focused on the Italian language and context.

Egomnia’s Emma page adds that Emma-6 also draws on more than 60,000 real prompts gathered through Emma-5. Egomnia has not published the new model’s context window.

A filter trained on the launch

Emma Guardrails 1.0 is a classification model that sits in front of the chatbot. It reads each prompt a user sends, sorts it into categories, and blocks those that fall into risk categories before they ever reach the language model.

The approach will be familiar to anyone in financial technology, where card fraud systems score a payment before it is authorised and learn from the attacks they have already seen. Egomnia built its filter on the same principle, from the prompts people actually sent during Emma-5’s short public life.

Testing in public, on purpose this time

Access is limited to beta testers for the first few days, and Egomnia’s site says thousands of users applied for the role. The company is asking them to probe the model with technical, professional, creative and complex questions, to look for its limits and errors, and to send feedback by email or through OpenEmmaLab, a community the company runs on Discord.

Emma-5 faced mass public testing from the day it launched. Emma-6 starts with a closed group of people who signed up specifically to find its flaws.

Where Emma-6 sits

Emma-6 remains small by the standards of the models Europe’s leading developers are releasing. Mistral Large 4, which the French company opened in public preview this month, lists 1.05 trillion parameters, close to 680 times as many. Italy’s larger sovereignty asset remains IT4LIA, the national supercomputer with more than 8,000 graphics processing units (GPUs).

Emma-5’s five days online left Egomnia with two things it could not have bought: more than 44,700 real prompts to train the filter that now screens Emma-6, and, by its own account, thousands of people asking to test the next version. That record of how Italians actually used an Italian model is, for a project built on Italian AI sovereignty, the one asset no foreign lab could have collected.

Editor’s note

Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. If you spot an error, have additional information, or want to flag anything, write to rosalia@thebrightminded.com.