---
title: README
emoji: 🧭
colorFrom: purple
colorTo: blue
sdk: static
pinned: false
---

![Vitund](vitund-header.jpg)

# Vitund

**Turning intents into action.**

We build the tools that let organizations put AI to work without giving up control.

**[Vitund Sandbox](https://vitund.ai/sandbox)** gives AI agents and coding assistants isolated Linux environments
where they can write code, call APIs and install packages, either self-hosted on your own infrastructure or as a
managed service. Each sandbox has its own network controls, credentials are injected at the network layer so agent
code never sees the keys, and every file, process and network call is recorded for audit.

**Cox** is our family of decision models. Give Cox a text and a set of questions with fixed answers (yes or no, one
of these options, a rating on this scale), and it returns a probability for every answer in a single pass, without
generating any text. Every answer is one of the options you offered, so the result drops straight into code, and the
probabilities are calibrated, so you can act automatically on the confident answers and send the rest to a person. Cox is named for Richard T. Cox (1898–1991), whose theorem (1946) showed that any consistent way of reasoning about uncertain statements must follow the rules of probability, which is why every answer Cox gives is a probability over the options you offered.

The two fit together, because an agent acting on someone's behalf keeps running into small decisions (may this
request leave the network, does this action follow the policy, is there enough information to go ahead) that need a
fast, checkable answer rather than another paragraph of generated text. We think a lot of the work teams hand to
large language models today is really a decision of this kind, and a decision model does it in tens of milliseconds
on one GPU, at a small fraction of the cost.

## Models

| model | size | licence | what it's for |
|---|---|---|---|
| [cox-1-3b](https://huggingface.co/Vitund/cox-1-3b) | 2.9B | Apache-2.0 | yes/no, choice and score decisions about English text, many questions per text |

Cox is fine-tuned from Qwen3.5-4B-Base on data only with licences that allow commercial use. The model card
lists every source. It reads text up to 64k tokens and answers each question independently, so one request can carry
as many questions about the same text as you need.

## Get started

```bash
pip install coxlm
```

```python
import coxlm
from coxlm import Questions, choice, yesno


class Ticket(Questions):
    team = choice(["billing", "support", "sales"], instructions="Which team should handle this?")
    refund = yesno("Is the customer asking for a refund?")


with coxlm.load("Vitund/cox-1-3b") as model:
    ans = model.decide("My card was charged twice for the same order.", Ticket)
    print(ans.team.choice, ans.team.confidence, ans.refund.p_yes)
```

`coxlm` is open source ([github.com/Vitund-AI/coxlm](https://github.com/Vitund-AI/coxlm)), and `coxlm-serve` runs the
model as an HTTP service, including a System One compatible endpoint.

## More

- Website and articles: [vitund.ai](https://vitund.ai)
- Contact: opensource@vitund.ai