---
pipeline_tag: text-to-speech
library_name: mlx-audio
base_model: rumik-ai/rumik-oss-1
base_model_relation: quantized
license: cc-by-nc-4.0
language:
- hi
- bn
- ta
- te
- mr
- gu
- kn
- ml
- pa
- or
- as
- ur
- ne
- sa
- mai
- mni
- brx
- doi
- kok
- sat
- ks
- en
tags:
- mlx
- tts
- text-to-speech
- indic
- expressive
- mimi
- speech
---
# rumik-oss 1 · MLX 4-bit
**4-bit** MLX quantization of [rumik-oss 1](https://huggingface.co/rumik-ai/rumik-oss-1), a 3B multilingual text-to-speech model for **22 Indic languages + English** with expressive delivery control and inline vocalizations, producing 24 kHz audio. Runs on Apple silicon with [mlx-audio](https://github.com/Blaizzy/mlx-audio).
Fastest. Noticeable fidelity loss (80% next-token agreement); treat as experimental.
## this export
| | |
|---|---|
| quantization | 4-bit affine, group size 64, stop predictor kept in bf16 |
| size on disk | 1.9 GB |
| peak memory | 2.1 GB |
| decode speed, M5 MacBook Air (16 GB) | 69 tokens/s (0.69x real time) |
| next-token agreement with bf16 | 80% |
100 audio tokens make one second of speech. The Mimi codec is downloaded automatically on first use.
## usage
```bash
pip install mlx-audio
```
```python
from mlx_audio.tts.utils import load
model = load("rumik-ai/rumik-oss-1-mlx-4bit")
for chunk in model.generate(
"नमस्ते, आज आपका दिन कैसा रहा?",
voice="Ira",
instruct="happy, Hindi accent, steady pace",
stream=True,
):
play(chunk.audio) # 24 kHz float32
```
```bash
python -m mlx_audio.tts.generate --model rumik-ai/rumik-oss-1-mlx-4bit \
--text "नमस्ते, आज आपका दिन कैसा रहा?" --voice Ira \
--instruct "happy, Hindi accent, steady pace" --stream
```
## controls
| control | values |
|---|---|
| voice | `Ira`, `Aisha`, `Siya`, `Zoya` |
| tone | happy, sad, angry, excited, professional |
| accent | Hindi, Telugu, Tamil, Kannada, Bengali, Punjabi, Indian English |
| pace | slow, fast, steady |
| inline | ``, ``, `` |
The description can also be written inline: ` जल्दी आओ! `
## samples
Samples above are from the original bf16 model; see the [original card](https://huggingface.co/rumik-ai/rumik-oss-1) for benchmarks (IndicEmo, NoVA, WER/CER) and limitations.
## license
Research and non-commercial use under [tiny aya fire's CC-BY-NC 4.0 with acceptable-use addendum](https://cohere.com/cohere-labs-cc-by-nc-license), unchanged from the original. `LICENSE` and `NOTICE` are included; this repo adds 4-bit weight quantization as a modification. The [Mimi codec](https://huggingface.co/kyutai/mimi) is CC-BY-4.0.
## citation
```bibtex
@unpublished{govindu2026rumikoss1,
title = {{rumik-oss 1 technical report}},
author = {Govindu Pranav and Anant Shukla and Suryansh Shakya and Aman Anand and Vatsal Bharti},
year = {2026},
note = {In preparation}
}
```
