Fararoni OS
The agent runtime that runs on your server, with a web interface, WhatsApp and Telegram. Early access from 3 November.
Learn more →Sovereign AI that works in production
SOVEREIGN AI INFRASTRUCTURE
Autonomous agents with deterministic RAG on infrastructure you control: from a 16 GB machine, designed to scale to NVIDIA GPU clusters. With local models, no sensitive data leaves your network.
Measured in September 2026
IS IT FOR YOU?
“I want to use AI with my documents and customer data, but I can't send them to a third-party cloud.”
If this sounds familiar, we can help.
Book a 20-minute demoNo commitment · We reply in less than 24 hours
The runtime reserves about 8 GB. Start with 16 GB: a fast text agent. Voice and long reasoning from 32 and 64 GB.
16 GB · CPU
Recommended minimum, one machine
32–64 GB · CPU
Department-wide operation
GPU server
High concurrency
NVIDIA cluster
Large models and mixture of experts
WHAT WE ARE BUILDING
Everything starts from the same principle: your information stays with you. On that core we build tools to work better.
The agent runtime that runs on your server, with a web interface, WhatsApp and Telegram. Early access from 3 November.
Learn more →A CRM that fits your size, and your budget.
SHARED
You pay for what you use. If nobody logs in, it shuts itself down.
DEDICATED
Your own instance, just for you. Nothing shared.
FARARONI OS · HOW TO START
OPEN SOURCE · APACHE 2.0
Available since March 2026 · published as C-FARARONI
Fararoni's original core: the agent engine in your terminal, with no graphical interface. For learning, experimenting and building on top.
EARLY ACCESS · 3 NOV
Limited spots · selection by application
Try the complete platform in your company before launch: web interface, WhatsApp and Telegram, with direct support from our team.
BUSINESS
For production
A version redesigned for production: hardened modules, with code and dependencies reviewed by an independent security provider (SAST and SCA), with support to install it on your network.
HOW IT WORKS
SOVEREIGN
Reasoning happens inside your network. You decide where models and data live.
CONNECTED
WhatsApp, Telegram, Discord and your internal systems, connected to the same agent.
GOVERNED
Every action is recorded, and a person approves what matters.
MODELS · DATA SOVEREIGNTY
You choose where each agent runs: on open models under your control, or with the cloud provider you already use. Mix them and switch anytime, without rewriting your agents.
FULL CONTROL · LOCAL
FOR SENSITIVE DATAOn your machine or server, even without internet, or in your private cloud. These models are already validated in real use; you can add other compatible ones, like a self-hosted Kimi, and assign them any role.
| MODEL | ROLE IN FARARONI | SIZE | FROM |
|---|---|---|---|
| Qwen3.5 9B | Rabbit · fast answers | 6.6 GB | 16 GB |
| Qwen3 30B Instruct | Voice · spoken conversation | 18.6 GB | 32 GB |
| Qwen3 4B Instruct | Summaries for voice | 2.5 GB | 32 GB |
| Qwen3.8 27B | Tortoise · long reasoning | 31.7 GB | 64 GB |
GROWS WITH YOUR HARDWARE
Via Ollama or another compatible server, such as NVIDIA NIM or vLLM. Estimated memory: 8 GB runtime plus the models in use.
FASTEST · CLOUD
WITH YOUR API KEYRunning in minutes, no hardware needed. Great for getting started or for tasks without sensitive data.
| PROVIDER | REFERENCE MODEL | STATUS |
|---|---|---|
| Anthropic Claude | Claude Sonnet 5 | ● Verified in real use |
| OpenAI | GPT-5.1 | ● Verified in real use |
| Google Gemini | Gemini 3.5 Flash | ● Verified in real use |
| DeepSeek | DeepSeek V4 Flash | ● Verified in real use |
| Kimi · Moonshot | Kimi K2.7 Code | ● Verified in real use |
| Qwen · Alibaba | ||
| Token Plan | Qwen3.7 Plus | ● Verified in real use |
Includes 12 models for agents. QWEN qwen3.6-flashqwen3.7-plusqwen3.7-maxqwen3.8-flashqwen3.8-maxqwen3.8-max-previewDEEPSEEK deepseek-v4-flash-0731deepseek-v4.1-flashdeepseek-v4-prodeepseek-v4-pro-0813GLM glm-5.2glm-5.3 | ||
| DashScope (API) | Qwen3.7 Plus | ○ Compatible |
| GLM · Zhipu | GLM-5.1 | ○ Compatible |
| Microsoft Foundry | Kimi K3 | ○ On request |
| Amazon Bedrock | Kimi K3 | ○ On request |
For companies already on Azure or AWS: Kimi K3 is contracted through your own account, without using Moonshot's API. We integrate it if your project needs it. | ||
Any OpenAI-compatible provider can be added without changing code.
Measured in September 2026. The brands mentioned belong to their respective owners; Fararoni is not affiliated with them.
LEARN
What it is, what it cannot do and why it sometimes makes things up. Explained in Spanish, without jargon.
NEXT STEP
In person in Veracruz–Boca del Río and Mexico City, or online from anywhere.
We show you Fararoni running offline, on hardware like yours.
DEMO · FREE
With our team and sample data. You see the AI answering with no connection.
ALREADY HAVE A USE CASE?
We define what to solve, with which hardware and how long it takes. You leave with a concrete proposal.
For pilot program companies, the assessment is free.
CONTACT
The same agent answers you on both channels. If you would rather talk to a person, write to us at contacto@fararoni.dev.


These are demo channels: messages go through WhatsApp and Telegram servers, so do not share confidential information. In your company, Fararoni runs inside your own network. Privacy notice.