---
library_name: mlx
license: mit
license_name: mit
base_model:
  - anyspeech/zipa-small-crctc-ns-no-diacritics-700k
tags:
  - automatic-speech-recognition
  - speech
  - ctc
  - ipa
  - mlx
  - quantized
  - bee
pipeline_tag: automatic-speech-recognition
---

# ZIPA Small CR-CTC NS No-Diacritics 700k (MLX Q8)

This repository contains an MLX-native Q8 checkpoint bundle for the ZIPA small CR-CTC non-streaming no-diacritics model.

## What This Is

- Base model family: `anyspeech/zipa-small-crctc-ns-no-diacritics-700k`
- Published repo target: `bearcove/zipa-small-crctc-ns-no-diacritics-700k-mlx-q8`
- Quantization target: MLX Q8
- Intended consumer: the Bee repository
- Reference implementation: `rust/bee-zipa-mlx` inside Bee

This is a Bee model bundle. It is not currently intended as a standalone general-purpose MLX package with its own separately published loader crate.

## Files

- `model.safetensors`: quantized MLX checkpoint in the `zipa-mlx-quantized-v1` format
- `tokens.txt`: CTC vocabulary
- `config.json`: model architecture and quantization metadata

## Quantization Scheme

- Linear layers are quantized to 8-bit where MLX supports the target group size
- Small incompatible projections remain dense inside the checkpoint
- Norms, bypass scales, convolution weights, and downsample weights remain dense
- Group size: `64`

## Usage From Bee

Example from a Bee checkout:

```bash
cargo run -q -p bee-zipa-mlx --bin zipa-infer -- \
  --quantized-checkpoint /path/to/model.safetensors \
  /path/to/audio.wav
```

## Notes

- This bundle was generated from local dense ZIPA reference artifacts and quantized with `bee-zipa-mlx`
- The artifact layout is currently project-specific to Bee
- The bundle is currently marked `MIT` based on the upstream ZIPA repository license and the upstream release statement that code and pretrained models would be released under permissive licenses. This should still be confirmed directly with upstream model authors.