How to Run UnimeType Transliteration-4B in LM Studio

UnimeType Transliteration-4B is a small Convert model available as Safetensors, GGUF, and MLX. This guide covers the MLX version in LM Studio on Apple silicon.

Download the model

For a hosted Pinyin-to-Chinese route, see the UnimeType transliteration API. Its Devnet preview exposes payment quotes; paid model calls are not open yet.

The three formats share the retained fine-tuned checkpoint. The standard export is not a dequantization of the published MLX 6-bit file. Fine-tuning used a quantized 4-bit base, so floating-point storage does not restore the original base model's pre-quantization precision. Training data is not public.

The September 6 checks used an Apple M5 Pro MacBook Pro with 48 GB memory. NVIDIA, AMD, iOS, and Android inference remain unverified, as does GGUF loading in LM Studio. Follow the GGUF model card for the tested llama.cpp command.

The standard and GGUF exports have a known Arabic trailing-newline error. Short English inputs and Japanese punctuation or emoji can also fail preservation checks. Model cards include actual test results, their limits, and file checksums; a download does not guarantee exact conversion on your input.

Open the Transliteration-4B model page and download the repository with your preferred Hugging Face workflow. In LM Studio, import or select the downloaded MLX model directory.

The published artifact is based on Qwen3.5-4B and uses 6-bit affine quantization with group size 64. The model repository is approximately 3.44 GB. Storage size does not equal peak runtime memory.

Load it as a text model

Use a context length of 2048 for the short Convert examples described here. Start the local server, then select the loaded model in LM Studio.

Convert expects a task prompt that identifies the target language and asks for replacement text only. A general chat prompt is not the same test and may produce commentary or a different writing task.

Try one input from each language

Useful smoke tests include:

These examples test more than script generation. Each one also checks whether an English span remains unchanged.

Test preservation before testing long prose

Before judging the model on paragraphs, try code, a product name, a URL, an emoji, and a multiline input. Use the model to convert Latin-character input into the target script. It is not intended for translation, rewriting, Polish, or Explain.

Local execution keeps inference on the machine running LM Studio. What another application sends to the local server depends on that application's integration and settings.

For the score behind the model card, read How we evaluate multilingual transliteration. The product's local and network boundaries are documented in Product boundaries.