ChindaMT-4B-MLX-4bit

ChindaMT-4B-MLX-4bit

This is an MLX-format conversion of iapp/ChindaMT-4B by iApp AI Research. All credit for the model weights, training, and translation capability belongs to the original authors. This repo only provides an Apple Silicon–compatible format. No retraining or modification of weights was done.

Model Details

Uses group_size=32 for finer quantization granularity. 94.6% semantic fidelity — best balance of size and quality for casual use.

Quick Start

pip install mlx-lm

mlx_lm.generate \
  --model veha/ChindaMT-4B-MLX-4bit \
  --prompt "Translate English to Thai.

EN: The meeting has been postponed until next Friday." \
  --max-tokens 256 --temp 0

Python:

from mlx_lm import load, generate

model, tokenizer = load("veha/ChindaMT-4B-MLX-4bit")
response = generate(model, tokenizer, "Translate English to Thai.\n\nEN: The weather is nice today.", temp=0)
print(response)

LM Studio

Download and load this MLX model in LM Studio:

lms get https://huggingface.co/veha/ChindaMT-4B-MLX-4bit --mlx -y
lms ls
lms load chindamt-4b-mlx@4bit --context-length 4096 --identifier chindamt-4b-4bit -y

Run a one-shot translation from the LM Studio CLI:

lms chat chindamt-4b-4bit \
  -p "Translate English to Thai.

EN: The meeting has been postponed until next Friday." \
  --stats -y

If LM Studio assigns a different model key on your machine, use lms ls and replace chindamt-4b-mlx@4bit with the listed key.

Prompting

Same prompt format as the original iapp/ChindaMT-4B:

Translate English to Thai.

EN: <source text>

Translate Thai to English.

TH: <source text>

With rules:

Translate English to Thai.
Rules:
- Return only the translated text
- Use a professional tone

EN: <source text>

MLX Variants

All converted from the same iapp/ChindaMT-4B weights:

*Speed measured on M3 Max 128GB.

Fidelity = cosine semantic similarity against original HuggingFace weights (multilingual sentence embeddings). 100/100 prompts identical for bf16, 81/100 for 8bit, 37/100 for 4bit.

Evaluation

Corrected 100-prompt evaluation, 50 EN->TH and 50 TH->EN pairs:

Evaluation script: eval/full_table.py. Metrics use multilingual sentence embeddings for semantic similarity, chrF++ with word_order=2, and character-level BLEU.

Conversion Method

This repo was built with MLX-LM 0.31.3 from the original Hugging Face model.

1. Download the original model after accepting access to iapp/ChindaMT-4B:

huggingface-cli download iapp/ChindaMT-4B \
  --local-dir hf/ChindaMT-4B \
  --local-dir-use-symlinks False

1. Convert and quantize the local Hugging Face weights to MLX 4-bit with group size 32:

mlx_lm.convert \
  --hf-path hf/ChindaMT-4B \
  --mlx-path mlx/ChindaMT-4B-q4 \
  --quantize --q-bits 4 --q-group-size 32

Note: --q-group-size 32 provides finer quantization granularity than the default 64.

1. Post-process the release folder:

• Keep Apache-2.0 license metadata and add a LICENSE file.
• Add .gitattributes so safetensors files are stored with Git LFS.
• Remove the forced <think> generation prefix from chat_template.jinja; this is a translation model, so LM Studio and other clients should display translations as normal assistant output, not reasoning text.
• Remove tool_parser_type: qwen3_coder from tokenizer_config.json; this model is released for translation, not tool calling.

1. Evaluate the quantized model against the original HF model and MLX bf16:

python eval/full_table.py

This Disclaimer, Translated by Every Variant

The disclaimer above, translated by each MLX variant. A real-world quality comparison on this very text:

EN→TH

TH→EN (round-trip: Thai back to English)

Quantization Comparison: Real Examples

How much does quantization actually affect translation quality? Here are hand-picked examples where the variants diverge:

Example 1: Proper Nouns (EN→TH)

Input: Christoph Müller will present at the ASEAN Business Summit in Bangkok on December 15, 2025.

Observation: FP and Q8 are identical. Q4 transliterates "Müller" as มูเลอร์ (Moo-ler) instead of มึลเลอร์ (Müller) — the umlaut nuance is lost.

Example 2: Context-Dependent Translation (TH→EN)

Input: ยอมแลกกับอะไรก็ได้ ("willing to trade for anything")

Observation: FP/Q8 try to explain the Thai phrase instead of simply translating it. Q4 produces a more direct translation — arguably better, but inconsistent with FP behavior.

Example 3: Tone & Gender Register (TH→EN)

Input: คุณครูคะ หนูขอถามหน่อยได้ไหมคะ ว่าการบ้านส่งวันไหนคะ (female student speaking to teacher, with polite particles คะ/หนู)

Observation: All variants handle honorifics correctly here — the polite register is preserved. This is a strength of the base model.

Example 4: Technical Terms (EN→TH)

Input: The WebSocket connection was reset by the proxy server. Please check your Kubernetes ingress configuration and retry.

Observation: FP and Q8 are identical. Q4 uses a slightly longer ending phrase but preserves all technical terms correctly.

Example 5: Long-Form News (TH→EN) — Where Q4 Really Diverges

Input:

การเปิดตัวรถไฟความเร็วสูงสายกรุงเทพฯ-นนทบุรีถือเป็นก้าวสำคัญของการคมนาคมขนส่งสาธารณะในประเทศไทย โครงการนี้มีมูลค่ากว่า 150,000 ล้านบาท ใช้เวลาก่อสร้างนานกว่า 8 ปี และคาดว่าจะสามารถรองรับผู้โดยสารได้วันละ 500,000 คน นอกจากจะช่วยลดปัญหาการจราจรติดขัดอย่างหนักบนถนนพหลโยธินแล้ว ยังเชื่อมต่อกับระบบรถไฟฟ้า BTS และ MRT ที่สถานีสำคัญหลายแห่ง ผู้ว่าการการรถไฟแห่งประเทศไทยกล่าวว่า แม้จะมีความท้าทายหลายประการในระหว่างการก่อสร้าง แต่ผลลัพธ์ที่ได้สะท้อนถึงความมุ่งมั่นของทีมวิศวกรและความร่วมมือจากภาครัฐและเอกชน ตั้งแต่วันเปิดให้บริการเป็นต้นไป ประชาชนสามารถใช้บริการฟรีตลอดทั้งเดือนแรก

Full output for reference:

Observation: FP and Q8 are perfectly identical on long text. Q4 stays accurate (correct budget figure) but drifts on word choices: "marks" vs "represents", "required" vs "has taken", "free service" vs "free access".

Takeaway

Recommendation: For professional use, 8-bit is the sweet spot (99.3% fidelity, nearly identical output, half the size). For casual use, 4-bit is fine.

Acknowledgements

• Original model: iapp/ChindaMT-4B by iApp AI Research
• Framework: MLX and MLX-LM by Apple