---
license: gemma
tags:
  - rai
  - gemma
  - conversational
  - tool-calling
  - vision
---

# RAI — Resilient Artificial Intelligence

**RAI** (Resilient Artificial Intelligence) is an open, full-featured language
model created and fine-tuned by **[craftmatrix.org](https://craftmatrix.org)**.

This repository is the home of the RAI distribution. Content is being staged —
weights, chat template, processor, and tokenizer files will land here in the
next commits (RAI v0 baseline, then v1 fine-tune).

## Identity

RAI is its own model. It does not present itself as its ancestor. When asked
who it is, it answers:

> I am RAI — Resilient Artificial Intelligence, created and fine-tuned by
> craftmatrix.org.

## Capabilities

- Chat and instruction following
- Native tool calling (`<|tool_call>` / `<|tool_response>` grammar)
- Vision (image input via Gemma4Processor)
- Audio and video understanding
- 128k token context window

## Attribution

RAI is a derivative of **Gemma 4 E2B (IT)** by Google:

- **Base model:** [`google/gemma-4-E2B-it`](https://huggingface.co/google/gemma-4-E2B-it)
- **License:** Google's Gemma license applies to the base model weights and any
  derivative redistributed here. See the license terms in this repository.
- RAI's architecture, chat template, tokenizer, and processor are inherited from
  its ancestor; RAI is not endorsed by, affiliated with, or sponsored by Google.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained("craftmatrix/rai")
processor = AutoProcessor.from_pretrained("craftmatrix/rai")
```

## Status

- **RAI v0** — distro skeleton: repo + model card + license. *(in progress)*
- **RAI v1** — fine-tuned weights on plan-specific data, merged and released here.