EraX-Translator-V1.0-mlx-bf16

MLX-VLM bfloat16 conversion of erax-ai/EraX-Translator-V1.0 for Apple Silicon.

Notes

• Converted locally with mlx-vlm 0.4.0.
• Source model architecture: Gemma3ForConditionalGeneration.
• This checkpoint is intended for translation tasks and was tested here on Vietnamese translation.
• Local smoke test passed with mlx-vlm text generation.
• Local evaluation was run on 5 Vietnamese translation cases covering English, German, modern Chinese, and Classical Chinese.

Conversion

python3 -m mlx_vlm convert \
  --hf-path erax-ai/EraX-Translator-V1.0 \
  --mlx-path /path/to/EraX-Translator-V1.0-mlx-bf16 \
  --dtype bfloat16

Quick Start

python3 - <<'PY'
from mlx_vlm import load
from mlx_vlm.generate import generate

model, processor = load('/path/to/EraX-Translator-V1.0-mlx-bf16')
messages = [
    {"role": "system", "content": "Bạn là trợ lý dịch thuật nhiều ngôn ngữ. Chỉ trả về bản dịch chính xác, không giải thích thêm."},
    {"role": "user", "content": "The weather is nice today, but the traffic is terrible.\n\nDịch sang tiếng Việt."},
]
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
result = generate(model, processor, prompt, verbose=False, max_tokens=128, temperature=0.2, top_p=0.95, top_k=64)
print(result.text)
PY

Validation Summary

• Case count: 5
• Avg generation speed: 37.68 tok/s
• Avg wall time: 4.62s
• Max peak memory: 10.42 GB
• Avg similarity to reference set: 0.2828
• Preamble leakage: 0 cases

Translation Quality

Observed locally:

• Strongest on everyday English and Chinese to Vietnamese translation.
• More stable than the 8bit variant on German prose and Classical Chinese Buddhist text.
• Better choice when name fidelity and wording stability matter more than speed.

Caution

This model is a translation-tuned checkpoint. It is not intended as a general-purpose coding or math model, and difficult literary or historical material may still require human review.