FRQ Get early access
Faruq-One · Foundation Decision Engine Apache 2.0

Decisions, not paragraphs.

The open foundation System One model for software and autonomous agents. Send a situation, a question, and your allowed choices. It returns the exact decision with calibrated probabilities in a single sub-20ms pass—zero tokens generated, zero hallucinations.

Get early access Hugging Face ↗
✓ Sub-20ms Latency ✓ 100% Apache 2.0 Permissive ✓ Calibrated Confidence

FRQ is Faruq without its vowels. Faruq, said fa-rook, is an Arabic name for the one who distinguishes. Its root, f-r-q, means to separate, to tell apart, to decide.

Context 
The user wants to find the cheapest roundtrip flights from New York to London departing Dec 12 and returning Dec 20.
Question

Which tool should execute this?

Decision
  1. search_flights96.4%
  2. book_hotel2.1%
  3. currency_convert1.0%
  4. check_weather0.5%
18.4 ms latency
0 output tokens
BFCL Tool Track
Calibrated T=3.0

Model

Most of what software asks an AI is a choice.

Which queue. Which tool. Yes or no. Safe or not. A System One model is built for that one job: it reads the question once and picks from the options it was given.

Typed in, typed out

You send a state, a question and the allowed options. You get back one of those options. Nothing else can come out.

One pass

Faruq is a 150M-parameter encoder. It scores every option in one read. No text is generated, so there is nothing to parse.

A number to act on

Every option comes back with a probability. Your code can accept the answer, check it again, or hand over to a person.

Type
Encoder
Size
150M parameters
Reads
Up to 8,192 tokens per request
Returns
One option, and a probability for each
Language
English. Arabic models come next.
Status
In development

Platforms

A decision step inside the tools you already use.

Faruq will come as nodes, apps and plugins that install where you already work, starting with automation platforms such as n8n, Make and Zapier. It reads what just arrived and tells the next step where to go. No code needed.

Trigger New support email "I was charged twice for my October invoice."
Faruq · decide Which team should handle this? Billing, technical support or sales
  • Billing94%Open a billing ticket
  • Technical support3%
  • Sales2%
Illustration. Nothing is released yet, and the list of platforms is not final.

  1. First

    Nodes, apps and plugins

    Installed from inside the platform you use. Pick the question, list the options, branch on the answer.

  2. Always

    The API

    The same model for your own software, from the first day.

  3. Later

    Companies and public institutions

    Working directly with larger teams that want Faruq inside their own systems.

You pay for decisions, not tokens. The plan is a price per decision, on its own or in bundles.

API

One request, several questions, one answer each.

Behind every workflow step is the same call. Describe the situation once, then ask as many questions about it as you need. Each question carries its own options.

You sendpreview
You get backpreview

A preview of the request shape. The API is not open yet and details may change.

  1. State

    Whatever the decision is about: a message, a document, a conversation so far, a list of tools.

  2. Questions

    Each one is an instruction plus the options to choose from, or a plain yes or no.

  3. Answers

    For each question, the chosen option and the probability Faruq gives to every option.

Uses

Where a fast, narrow answer is the right one.

Each use below is paired with the public benchmark that tests it, and what Faruq scores on it today.

Route a message

Read a customer's message and send it to the right intent, queue or team.

Decide if a tool is needed

For each tool an agent has, say whether this request calls for it.

Choose the next step

Call a tool, ask a follow-up question, answer directly, or decline.

Check a contract

Does the agreement support a statement, contradict it, or not mention it?

Read sentiment about a company

How a piece of financial news feels about each company it names.

Find the right tool among many

Rate a long list of candidate tools by how useful each is for a request.

What it is not for

Maths, graduate-level science, chess and long chains of reasoning. A small model that answers in one pass does not do these, and Faruq scores at or near chance on them today. Use a larger model there.

Results

Measured, not promised.

Numbers from the current development build. Faruq is still in training, so they will change.

–

Index on the public Decision Index

–

Right on tasks it never trained on

Skill above chance, by benchmark

0 is guessing, 100 is perfect. marks a benchmark whose public training set Faruq trained on.

    How to read these numbers

    • The index is our own run on a fixed sample of the public benchmarks of the Decision Index. It is not an official leaderboard entry.
    • The leaderboard's main score also counts private tests that only its maintainers can run. Small models usually score lower there than on the public part.
    • A request that is too long for Faruq is refused and counted as wrong. Nothing is cut short to make it fit.

    Early access

    Try Faruq before it opens.

    Leave your email and tell us which decision you would hand to Faruq. We will write to you when there is something to try.

    Where would you use it? optional

    We use your email only to contact you about Faruq.