metadata
language:
- en
- sw
- ha
- yo
- ig
- am
- zu
- xh
- af
- so
- rw
- sn
- tw
- ee
- wo
- ny
- ti
- nso
- tn
- om
- ve
- nd
- ar
- fr
- pt
- es
- de
- zh
- ja
- ko
license: mit
tags:
- translation
- mlx
- apple-silicon
- multilingual
- african-languages
pipeline_tag: translation
library_name: mlx
TranslateBlue v2 (MLX 4-bit)
Translation model focused on 29 languages with emphasis on African languages, in MLX 4-bit format for Apple Silicon (M1+ Mac, and mlx-swift on supported devices).
Model description
- Base model: Qwen3-4B-Instruct
- Format: MLX, 4-bit quantized
- Size: ~2.1 GB
- Training: LoRA fine-tuning on parallel translation data (10,000 steps, 16 LoRA layers)
- Training data: 563,986 sentence pairs from 29 languages
Intended use
- Text translation between the supported languages, especially to/from African languages
- Offline translation on Mac (and in apps using mlx-swift where supported)
- Low-latency translation on Apple Silicon with Metal acceleration
Supported languages (29)
| Code | Language | Code | Language | Code | Language |
|---|---|---|---|---|---|
| sw | Swahili | ha | Hausa | yo | Yoruba |
| ig | Igbo | am | Amharic | zu | Zulu |
| xh | Xhosa | af | Afrikaans | so | Somali |
| rw | Kinyarwanda | sn | Shona | tw | Twi |
| ee | Ewe | wo | Wolof | ny | Chichewa |
| ti | Tigrinya | nso | Northern Sotho | tn | Tswana |
| om | Oromo | ve | Venda | nd | Ndebele |
| ar | Arabic | fr | French | pt | Portuguese |
| es | Spanish | de | German | zh | Chinese |
| ja | Japanese | ko | Korean | en | English |
Limitations
- Apple Silicon only for this MLX build (Mac with M1 or later; mlx-swift on supported iOS/iPadOS when available).
- Best for short to medium sentences; very long texts may lose quality.
- Low-resource pairs may be less accurate than high-resource ones.
- No built-in language detection; source and target languages should be specified in the prompt.
How to use
Prompt format
Use a clear translation instruction, for example:
Translate from English to Swahili:
Hello, how are you?
With Python (mlx-lm)
pip install mlx mlx-lm
from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler
model, tokenizer = load("aoiandroid/TranslateBlue-v2-MLX-4bit")
sampler = make_sampler(temp=0.3, top_p=0.9)
messages = [{"role": "user", "content": "Translate from English to Swahili:\n\nHello, how are you?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=64, sampler=sampler, verbose=False)
print(response)
With Swift (mlx-swift / TranslateBlue)
The model is registered as TranslateBlue v2 (MLX). After downloading via the app (or placing the model in the expected path), it runs with MLXModelService using the same prompt format above.
Training details
| Setting | Value |
|---|---|
| Base model | Qwen3-4B-Instruct |
| Method | LoRA |
| LoRA layers | 16 |
| Steps | 10,000 |
| Training samples | 563,986 |
| Validation loss | ~2.5 |
Related models
- GGUF version (llama.cpp, cross-platform): aoiandroid/TranslateBlue-v2-GGUF
License
MIT.
Citation
If you use this model in research or a product, please cite the base model (Qwen3) and the TranslateBlue project as appropriate.