---
title: Borealis Inference
emoji: 🎙️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.9.1
app_file: app.py
pinned: false
license: apache-2.0
models:
  - Vikhrmodels/Borealis-5b-it
---

# Borealis-5B-IT Inference

Audio-Language Model for Speech Understanding.

## Features

- Upload audio or record from microphone
- Multiple prompt presets (transcription, summarization, Q&A)
- Support for Russian and English
- Customizable generation parameters

## Model

- **Architecture**: Whisper Large V3 (encoder) + Qwen3-4B (LLM)
- **Parameters**: ~5B
- **Languages**: Russian, English

## Usage

1. Upload an audio file or record using microphone
2. Select a prompt preset or write custom prompts
3. Adjust generation parameters if needed
4. Click "Generate" to get the response

**Note**: Running on CPU, generation may take some time.

## Links

- [Model Card](https://huggingface.co/Vikhrmodels/Borealis-5b-it)
- [Training Datasets](https://huggingface.co/datasets/Vikhrmodels/Speech-Instructions)