---
license: other
license_name: gemma
license_link: https://ai.google.dev/gemma/terms
tags:
  - coreml
  - apple-neural-engine
  - ane
  - gemma
  - gemma-3
  - on-device
base_model: google/gemma-3-4b-it
---

# Brevox — Gemma 3 4B IT (ANE) Asset

Archived Core ML bundle used by **[Brevox](https://github.com/teleng-labs/brevox-ios)**
as the on-device fallback summarizer when Apple FoundationModels rejects a
given recording via its non-bypassable safety guardrails.

This repository **only hosts the distribution artifact** (the `.aar` file) for
the Brevox iOS/macOS app's runtime download. It is not the source of the
conversion pipeline nor a redistribution of upstream Google weights in their
original form.

## What is this?

A single Apple Archive (`.aar`, LZFSE-compressed) containing a Core ML 3-stage
runner of Gemma 3 4B IT (int4 QAT, 4096 context) converted by
[ANEMLL](https://github.com/anemll/anemll). Extracted at runtime into
`~/Library/Application Support/Brevox/Models/Summarizer_Gemma3_4B/`.

| Field | Value |
|---|---|
| File | `Summarizer_Gemma3_4B.aar` |
| Size | ~3.85 GB (3,940 MB) |
| SHA-256 | `9fcd4c09ca3d66d3d15e60c8a2f212deabf648489d2d41f9d9a7ab3e6aef7fb1` |
| Compression | Apple Archive, LZFSE |
| Contents | 5 `.mlmodelc` (embeddings + 3 FFN chunks + lm_head) + tokenizer.json + tokenizer.model + config.json + meta.yaml |
| Converted by | [ANEMLL v0.3.5](https://huggingface.co/anemll/anemll-google-gemma-3-4b-it-qat-int4-unquantized-ctx4096) |
| Target hardware | Apple A17 Pro / M-series with Neural Engine |

## Provenance

1. Upstream weights: [`google/gemma-3-4b-it`](https://huggingface.co/google/gemma-3-4b-it) (QAT int4-unquantized variant).
2. Conversion to Core ML by ANEMLL:
   [`anemll/anemll-google-gemma-3-4b-it-qat-int4-unquantized-ctx4096`](https://huggingface.co/anemll/anemll-google-gemma-3-4b-it-qat-int4-unquantized-ctx4096).
3. Packaged into a single LZFSE Apple Archive for delivery via Brevox's
   `ModelDeliveryService` (`URLSession` + Range resume + streaming SHA-256 +
   AppleArchive extraction).

## Licensing

- Weights: governed by the **[Gemma Terms of Use](https://ai.google.dev/gemma/terms)** and the
  [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy). By
  downloading or using this asset you agree to those terms.
- Conversion pipeline: ANEMLL, MIT License.
- Packaging wrapper (this repo's `.aar`): MIT, trivially derivative.

## Why Hugging Face and not GitHub Releases

GitHub Releases caps individual asset size at 2 GB. The packaged bundle is
3.85 GB. Hugging Face's LFS-backed storage and anonymous CDN distribution are
the appropriate hosting choice.

## Usage (app side)

Brevox fetches the asset at:

```
https://huggingface.co/robertteleng/brevox-gemma-3-4b-ane/resolve/main/Summarizer_Gemma3_4B.aar
```

Manual verification:

```bash
shasum -a 256 Summarizer_Gemma3_4B.aar
# expected: 9fcd4c09ca3d66d3d15e60c8a2f212deabf648489d2d41f9d9a7ab3e6aef7fb1
```