News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Velvet AI

Velvet AI

Italian multilingual open-weight LLMs built for energy efficiency

Velvet AI is a family of multilingual generative language models from Almawave, an Italian AI company. Released from 29 January 2025, the open-weight models were trained on the Leonardo supercomputer and are built for energy-efficient use across cloud, on-premise and edge deployments.

Share on WhatsApp

  • Built in Italy by Almawave and trained on the Leonardo supercomputer at CINECA
  • Velvet 14B trained on over 4 trillion tokens across six languages
  • Italian makes up about 23% of Velvet 14B's training data
  • Over 400 billion code tokens from more than 100 programming languages
  • Released as open weights under Apache-2.0 on 29 January 2025

What it is

Velvet AI is a family of multilingual generative language models made by Almawave, an Italian AI company. The line includes Velvet 2B, Velvet 14B and a Velvet 25B variant. These are foundational large language models, designed and built in Italy on Almawave's own architecture, with a strong emphasis on Italian alongside other languages. They were trained on the Leonardo supercomputer operated by CINECA and published as open weights under Apache-2.0. Almawave positions them as energy-efficient systems that fit European regulatory frameworks, and they are ready for cloud, on-premise and edge use.

How it came about

Almawave, an Italian AI company, developed the Velvet family as part of a strategic push into high-performance, energy-efficient AI aligned with European regulation. Training ran on the Leonardo supercomputer, which CINECA manages. The models were released on 29 January 2025 in open-weight form. Velvet 2B and Velvet 14B are the documented members of the family, and a Velvet 25B variant is also listed. The models are integrated into Almawave's AIWave platform.

Inside the model

Velvet 2B carries 2 billion parameters and was trained in Italian and English on more than 2 trillion tokens. Velvet 14B saw over 4 trillion tokens spanning six languages, with Italian making up roughly 23% of that mix. It also absorbed over 400 billion tokens drawn from more than 100 programming languages, which Almawave says supports more structured reasoning. The models rest on Almawave's proprietary architecture and were trained on the Leonardo supercomputer at CINECA. Energy efficiency is a stated design goal throughout.

What it can do

The models handle six languages, with Italian given particular weight. Velvet 14B's training mix includes code from over 100 programming languages, intended to help it produce more structured inferences. Velvet 2B is narrower, covering Italian and English. Almawave describes the family as energy-efficient and suited to European regulatory requirements. No benchmark scores are reported here.

How to use it

All Velvet models are released as open weights under the Apache-2.0 licence. They are prepared for deployment in the cloud, on-premise and at the edge, and are built into Almawave's AIWave platform. The models run as a web app and support Italian, English, Spanish, Brazilian Portuguese, German and French.

Safety and testing

Almawave frames the Velvet family as designed to align with European regulatory frameworks. No further safety evaluations or incidents are described here.

Why it matters

The Velvet family represents Almawave's investment in high-performance, energy-efficient AI built in Italy and oriented towards European rules. The models are integrated into the company's AIWave platform and are prepared for use across cloud, on-premise and edge settings.

Questions people ask about Velvet AI

What is Velvet AI?

It is a family of multilingual generative language models created by Almawave, an Italian AI company. The line includes Velvet 2B, Velvet 14B and a Velvet 25B variant. They are foundational large language models built in Italy, released as open weights and aimed at energy-efficient operation.

How large are the models?

Velvet 2B has 2 billion parameters and was trained on over 2 trillion tokens in Italian and English. Velvet 14B has 14 billion parameters and saw over 4 trillion tokens across six languages. A Velvet 25B variant is also listed as part of the family.

What languages do they support?

The models cover Italian, English, Spanish, Brazilian Portuguese, German and French, with Italian given particular emphasis. Velvet 14B's training data is about 23% Italian, and it also includes code from more than 100 programming languages to help with structured inferences.

How can I use them?

The weights are published under the Apache-2.0 licence, so they can be downloaded and run freely. Almawave says the models are ready for cloud, on-premise and edge deployment, and they are integrated into the company's AIWave platform.

What were they trained on?

Training ran on the Leonardo supercomputer managed by CINECA. Velvet 14B used over 4 trillion tokens across six languages plus over 400 billion tokens of code, while Velvet 2B used over 2 trillion tokens in Italian and English.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Velvet AI", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.