| --- |
| license: apache-2.0 |
| base_model: NiuTrans/LMT-60-4B |
| library_name: coreai |
| pipeline_tag: translation |
| tags: |
| - coreai |
| - ios |
| - on-device |
| - translation |
| --- |
| |
| # LMT-60-4B for Core AI on iOS |
|
|
| This repository contains an iOS 27 Core AI conversion of |
| [`NiuTrans/LMT-60-4B`](https://huggingface.co/NiuTrans/LMT-60-4B), pinned at |
| revision `f3d1043332179618de226eab7dd38c7290a7dfb9`. |
|
|
| It was exported with Apple's |
| [`coreai-models`](https://github.com/apple/coreai-models) toolchain at revision |
| `917dc99f0cc23ac0905c4de26e18e5a3764703fc` using the official Qwen3 4B iOS |
| mixed 4-bit/8-bit compression recipe. The maximum context length is reduced to |
| 2,048 tokens for live translation, lowering KV-cache pressure while retaining |
| ample context for short speech transcript segments. |
|
|
| ## Artifact |
|
|
| - Model: Qwen3-based LMT-60-4B |
| - Platform: iOS 27 Core AI |
| - Weights: mixed 4-bit/8-bit palettization (`qwen3_4b_mixed_4bit_8bit`) |
| - Context: 2,048 tokens |
| - Tokenizer: embedded |
| - Asset format: architecture-neutral `.aimodel` |
|
|
| The source `.aimodel` is specialized by Core AI for the user's device. This |
| keeps one downloadable artifact compatible with supported Apple Intelligence |
| devices instead of storing a separate multi-gigabyte specialization per device |
| architecture. |
|
|
| ## Languages |
|
|
| LMT-60-4B supports 60 source languages, including Croatian and Italian. It does |
| not include Slovenian; applications should use a dedicated Slovenian fallback. |
|
|
| ## License |
|
|
| The source model is released under Apache-2.0. Apple's conversion/runtime source |
| is governed by the BSD-3-Clause license in the upstream `coreai-models` |
| repository. |
|
|
