SKYL4R
Request a demo

Made in Italy. Runs everywhere.

Skylar is our foundation model, built from scratch. Skylar Web deploys it on your infrastructure.

Request a demoHow we train it
SkylarSkylar model
Ask Skylar…

Your AI agent, on your own servers.

Scroll to see what Skylar does every day, inside Skylar Web.

  1. Ask your documents

    Attach a contract, a policy, a manual. Skylar answers from the text and cites the passage it used, so anyone can check it.

    SkylarSkylar model
    Ask Skylar…

  2. When the answer isn't there, it says so

    If the document doesn't hold the answer, Skylar tells you instead of making one up. Then you know what is missing.

    SkylarSkylar model
    Ask Skylar…

  3. It uses your company's tools

    With function calling Skylar calls the tools you connect to it, like a ticketing system or an internal registry, and tells you what it did.

    SkylarSkylar model
    Ask Skylar…

  4. Compare models side by side

    In Skylar Web you can ask two models the same question, hide their names, vote for the better answer and export the report. You choose with evidence.

    SkylarSkylar model
    Blind comparisonExport report
    Model A
    Model B
    A is betterTieB is better

What it can do, out of the box.

The nine examples on the home page of Skylar Web. The last three call demo tools with sample data: in your company, Skylar calls the tools you connect.

  • Answers from your document

    A policy, a contract, an email: it answers from the text and cites the source.

  • Says when it's not there

    If the document doesn't hold the answer, it tells you instead of making one up.

  • Writes emails and texts

    A polite reminder to a client, ready to send.

  • Translates it ⇄ en

    A message in the other language, with the right tone.

  • Extracts data as JSON

    Name, email and city out of free text, ready for an app.

  • A COBOL fragment

    Adds up a table of amounts in a paragraph that compiles.

  • Checks an order

    Calls the right tool and answers with carrier, tracking and arrival date.

  • Weather and currency, together

    Two tools in one answer; tomorrow's date it works out by itself.

  • Books a meeting

    Fills title, date, time, length and attendees for the calendar.

Why Skylar, and not an AI through an API.

Many AI products send your documents to someone else's model, in someone else's cloud. We built the whole stack ourselves, from scratch, so it runs where your data already is.

SkylarAI through an API
Where it runsOn your server, next to your systemsIn the provider's cloud
Your documents and conversationsStay on your serverTravel to the provider with every request
The modelOurs, trained from scratch, no third-party weightsSomeone else's, rented by the token
Without internetKeeps workingStops answering
Who decidesYou, with an Italian startup that builds its own modelsThe provider: model, prices and terms
What it costsA plan per team, with a token budgetA bill per million tokens that grows with use

Sovereign by design.

Skylar Web is a single Docker container on a Linux server inside your perimeter. It sits next to your mainframe, not on it.

Your perimeter
Mainframe and legacy systems
Skylar Web server
Linux, Docker, CPU or NVIDIA GPU
Documents and conversations
SKYL4R portal
Sign-in, licence check, token count
Hugging Face
Model weights, first start only

Documents and conversations stay on your server. Only sign-in, the licence check and the plan's token count go to the SKYL4R portal: numbers, never text.

Next to the mainframe

It installs on a Linux server beside your systems. Nothing to install on the mainframe itself.

CPU or GPU

One Docker image for CPU and one for NVIDIA GPU. Start on the hardware you have.

Works offline

The licence is a signed permit that renews by itself: Skylar keeps working when the connection drops.

Plans sized by people.

Access with your SKYL4R customer portal account. Every plan has an annual token budget and a monthly cap.

up to
3
people
up to
10
people
up to
50
people

Prices on request.

Ask for a quote

Models we train ourselves, from zero.

No third-party weights under Skylar. Data, tokenizer, pre-training, post-training and evaluation are all done in-house.

01Data02Tokenizer03Pre-train04Post-train05Evaluate06Ship
01 / 06

Data

Data we collect, clean and curate ourselves. The more the model reads, the more it learns.

Pre-trained on Leonardo

174.3Btokens of pre-training

Our 990M base model was pre-trained from scratch on Leonardo BOOSTER, hosted by CINECA in Italy, with resources awarded by EuroHPC JU.

We acknowledge EuroHPC JU for awarding the project ID EHPC-AIF-2026PG01-1492 access to resources on Leonardo BOOSTER hosted by CINECA, Italy.

About EuroHPC AI Factories

Every model has to beat the previous one

Each training run ends with the same suite, new model against old, followed by a blind comparison.

  • COBOL that has to compile
  • 877 questions answerable only from a document
  • Instruction following
  • Tool use
  • Long context
  • Hallucinations

COBOLEval: a 980M model ahead of the 7B ones

146 problems, compiled and executed by GnuCOBOL, same harness for every model. Skylar-980M-Cobol has about 7 times fewer parameters. It is a research preview and a COBOL specialist, not a general chatbot.

Compiles
Skylar-980M-Cobol80.1%
StarCoder2-7B48.6%
CodeLlama-7B6.8%
Qwen2.5-Coder-7B6.2%
pass@1
Skylar-980M-Cobol7.5%
StarCoder2-7B0.0%
CodeLlama-7B0.7%
Qwen2.5-Coder-7B2.1%
Numbers from the model card

Our models, public and in training.

Part of our research is public: models on Hugging Face under Apache 2.0, and the framework that trains them on GitHub. The newest ones come out with their model card.

What we are training now

Not public yet: each one comes out with its model card.

  • Beta

    Skylar2-990M-Instruct

    The new Skylar 2 architecture, instruction-tuned for documents and tools.

  • In development

    Skylar2-990M-Cobol

    The COBOL specialist on Skylar 2, for an agent that works on a mainframe.

Public on Hugging Face

1,495 downloads across our 7 public models (Hugging Face, 4 November 2026).

All models on Hugging FaceFramework on GitHub

We built it at 1B. Now we take it to 32B.

SKYL4R is an early-stage Italian startup. Everything under Skylar is ours: the data pipeline, the tokenizer, the training stack, the architecture and the product companies install. We have proven the whole chain at 1B. Scroll to see what the next sizes take.

You are here · 1B
1Bparameters
×1
compute, against our 1B run
investment needed

Today: 1B

The whole chain, from data to Skylar Web, built by us and proven on a 1B model. This is where we are.

1B4B8B14B32B
You are herewith investment →

The multiplier is the compute of our 1B run: at the same tokens per parameter, compute grows with the square of the size. Every size from here on already exists in our framework as a Skylar 2 hybrid.

Invest in SKYL4R
  • Pre-training compute awarded by EuroHPC JU on Leonardo BOOSTER
  • 7 open models on Hugging Face, 1,495 downloads
  • A 980M model ahead of 7B ones on COBOLEval

See Skylar on your documents.

Write to us: we'll show it to you on documents like yours and tell you what it needs to run in your company.

Request a demo