Velvet: Shaping the future of AI through ethics and innovation

2025 will be a pivotal year for artificial intelligence, marked by consolidation and, most importantly, specialization.

AI is now a global focus, with more players beyond just U.S. and Chinese tech giants.

Vertical and specialized large language models (LLMs) are emerging all around the world, with models that don’t just understand language but also capture its cultural nuances, evolution, and context across different countries.

Velvet fits perfectly into this category. Developed by Almawave, this entirely Italian-made family of LLMs is lightweight, agile, and reliable, designed for both public and private organizations looking to harness AI’s potential with a strong focus on ethics, sustainability, and trustworthiness.

What sets Velvet apart is its ability to combine power and precision with a sleek and efficient architecture, optimized specifically for the Italian language and mindful of European ethical and legal regulations.

The two models that make up the Velvet family—14B and 2B—were fully developed by Almawave in Italy, built on proprietary architecture, and trained on the Leonardo supercomputer, managed by Cineca—one of the most powerful computers in Europe and the world.

With years of expertise gained through collaboration with businesses and public administrations, Almawave designed Velvet to offer an AI that is truly aware and contextualized.

In this article, we’ll explore all the key aspects of Velvet, its advantages, and its specific use cases.

As we’ve seen, Velvet is a family of multilingual Large Language Models designed and developed in Italy, built on strong data governance foundations and crafted with particular attention to the European regulatory framework.

These are advanced text-based AI tools capable of handling a wide range of linguistic tasks with high precision. For example, they can answer questions, provide machine translation, extract key information, automatically categorize texts and documents, and generate content enriched with access to external data sources.

Velvet is made up of two models, both built from scratch, designed to meet different scalability and specialization needs. What distinguishes them is primarily the number of parameters used, which indicates the complexity of the AI’s neural network and its ability to process information and generate content.

Velvet-14B

Velvet-14B is designed to handle complex tasks that require greater processing power.

Technical features:

• 14 billion parameters
• Trained with 4 trillion tokens and a vocabulary of 127,000 tokens—the equivalent of 50 million books
• Training data was algorithmically processed, reducing toxicity by 60% compared to the initial open-source data, ensuring greater control over response quality
• The model was refined with 50,000 safety instructions and trained on 2 million real-world examples from fields and sectors where Almawave has extensive experience
• Context window of 128,000 token

Velvet-2B

Velvet-2B is lighter and designed for more streamlined applications while still maintaining a high level of effectiveness.

Technical features:

●        2 billion parameters

●        Trained with 3 trillion tokens in 2 languages (Italian and English)

●        1 million real-world examples

●        Context window of 32,000 tokens

What makes Velvet truly innovative is its approach, based on three key principles: awareness, lightness, and agility.

• Awareness: Velvet is designed with a particular focus on ethics and safety. The use of clean, controlled data, with an emphasis on reducing bias and toxicity, and strong attention to European regulations, makes it “aware” of the context in which it operates, especially concerning data protection and the reliability of responses.
• Lightness: Velvet is designed to be efficient and sustainable. The Velvet-2B version can run on less powerful infrastructures without requiring massive computational resources. This makes it accessible even to smaller organizations. Energy efficiency is another key aspect—the optimized infrastructure reduces the carbon footprint and computational costs, ensuring a sustainable use of resources.
• Agility: Velvet is an AI designed to transform businesses and simplify people’s experiences in a practical way. Built to easily adapt to specific fields and industry languages, it can connect to internal processes and data within companies, integrating seamlessly into specialized and “ready-to-use” applications.

Velvet is designed for flexibility and versatility, making it easy to embed into specific applications and adapt to various tasks. Most importantly, it can be quickly customized for a wide range of industries and organizations, both public and private.

An ethical and technical committee oversees Velvet to guarantee it aligns with the values of transparency, fairness, and safety. Periodic reviews of Velvet assess the impact of the model in high-risk applications.

But this isn’t the only way the Almawave team has ensured adherence to these principles:

• Bias reduction and cultural context respect: During Velvet’s training process, specific methods were developed and integrated to minimize biases related to culture, ethics, hate speech, and gender. Additionally, unlike models primarily trained on English content or translated texts, Velvet’s training dataset was carefully balanced across multiple languages. For example, in Velvet-14B, 23% of the data consists of content originally written in Italian. This approach ensures that the generated results more accurately reflect the cultural differences and nuances inherent to the languages represented in the training data.
• Focus on privacy and European regulations: Velvet was trained using only controlled, clean, and reliable data to minimize toxicity and errors. To maximize privacy protection, an innovative feature of Velvet in this regard is “PAE” (Privacy Association Editing), a proprietary algorithm that allows for the removal of sensitive information directly from the model, when necessary, without the need for retraining.
• Not classified as a systemic risk model: Thanks to its balanced training architecture and model design, Velvet does not fall into the category of systemic risk models according to the AI Act definition, relieving distributing organizations from significant management responsibilities.
• Open-weights format under Apache 2.0 license, available in the cloud or on-premises: The models are available in an open-weights format, meaning that the model weights (the parameters resulting from the AI’s training) are public and transparent. The Apache 2.0 license is open-source and allows the use, modification, and distribution of the model and its weights for commercial purposes, as long as the original author is credited, and the license is included. Additionally, the models are available both in the cloud (on remote servers accessible via the internet) and on-premises (installed directly on the company or organization’s servers).
• Excellence partners: The Velvet models are also the result of numerous ongoing collaborations with the academic and research world to enhance their reliability. Partners include the University of Tor Vergata, the Bruno Kessler Foundation, La Sapienza University, the University of Catania, and the University of Bari. Additionally, SIpEIA (Italian Society for Ethics in Artificial Intelligence) has been involved for ethical and regulatory compliance.

One of the distinguishing features of Velvet, as we’ve seen, is its computational efficiency, achieved through a highly optimized dataset process and reduced toxicity.

Both the 14B and 2B models share a kind of data economy that does not compromise the quality of the outputs.

The goal is to create a super-efficient AI capable of seamlessly adapting to numerous vertical contexts.

How was this goal achieved?

• Training on real-world examples: To ensure rapid adaptation to various vertical contexts and greater accuracy in specific tasks, Velvet has been trained using concrete examples, based on Almawave’s years of experience in sectors such as healthcare, public administration, security, finance, mobility, education, and tourism. Specifically, Velvet-14B was trained on 2 million real-world examples, while Velvet-2B was trained on 1 million examples, all manually selected and drawn from real-life use cases.
• The context window size: The context window refers to the maximum number of tokens the model can process in a single input. In the case of Velvet-14B, the window can reach up to 128,000 tokens, equivalent to over 400 pages of text. The larger the context window, the more complex the requests made to the model can be. This means Velvet can handle even highly intricate tasks. This feature is crucial for making the model perform well in highly specific fields.

Key sector applications of Velvet

Velvet can be applied in a wide range of fields, thanks to its specialization and ability to process texts with connections, synthesis, and advanced reasoning.

Here are some practical examples of its applications:

• Healthcare: Decision support in DSS (Decision Support System) to assist medical teams in conducting quick and accurate analyses
• Public Administration: Providing personalized services to citizens, processing administrative documents, conducting legal and compliance analysis, streamlining internal operations
• Education: Advanced and personalized virtual assistants for students and teachers, both in corporate settings and schools
• Finance: Offering personalized consulting and management to users on banking platforms
• Justice: Support in document analysis and management of legal cases to increase transparency and accessibility, and reduce waiting times
• Security: Predictive analysis and advanced monitoring tools
• Mobility: Optimization of public transport information, advanced field intervention management, and user assistance
• Industrial and operational sector: Support in field operations through intelligent voice commands
• Customer service and contact centers: Advanced chatbots to improve user experience and operational efficiency
• Tourism: Automatic data analysis and review management for destinations and operators; personalization of the travel experience for tourists.

To facilitate the use of artificial intelligence across various sectors, Velvet has been designed to natively integrate with AIWave, Almawave’s PaaS (Platform as a Service) platform, which offers a wide range of ready-to-use AI solutions for different industries.

What is AIWave?

AIWave is a modular platform designed to facilitate the adoption of AI in business processes.

Notably, AIWave ensures a low-code/no-code approach, allowing even users without technical expertise to implement AI solutions. The platform equips users with technologies, tools, models, and features to seamlessly incorporate AI and natural language processing into existing business applications or create new ones that harness the power of language analysis and processing.

The integration of Velvet into AIWave strengthens the platform, amplifying its power and improving companies’ capacity to automate and optimize their processes.

The future of artificial intelligence in Europe

Velvet goes beyond being just an alternative to AI models from global tech giants. It introduces a groundbreaking paradigm based on specialization, ethics, and transparency. This empowers businesses and institutions with greater independence, reducing reliance on foreign solutions, while ensuring full compliance with local regulations.

Curious to learn more and try Velvet?

Contact us.