yacht/ChindaGo-4B-MLX-4bit

Model Summary

ChindaGo-4B is a token-efficient language model fine-tuned for Thai and English daily life tasks.
It is designed to handle text and images modalities, while generating concise, resource-aware answers suitable for on-device and edge deployment.

This model powers the ChindaGo Application, a practical AI assistant for iOS and edge devices, enabling offline, energy-efficient, and multimodal interactions for everyday life and emergency use cases.

• Base Model: google/gemma-3-4b-it
• Languages: Thai, English (multilingual support)
• Capabilities: Text and multimodal understanding (images, vision inputs)
• Optimized for: iOS, Android, and edge devices (offline, low-power environments)
• Domains: Daily life, Q&A, emergencies, personal assistance
• License: Apache 2.0

MLX Conversion

This model yacht/ChindaGo-4B-MLX-4bit was converted to MLX format from yacht/ChindaGo-4B using mlx-lm v0.27.1, with 4-bit quantization for memory efficiency on Apple Silicon devices.

Design Philosophy

ChindaGo-4B was fine-tuned to prioritize practicality over verbosity:

• Concise responses: tuned to avoid long, token-heavy outputs
• Resource-aware: optimized for mobile/edge inference with limited compute and battery
• Multimodal readiness: text + image
• Cultural alignment: adjusted for Thai context and daily-life interactions

Key Features

• Optimized for edge and mobile devices, enabling efficient on-device inference
• Token-efficient design for concise, low-latency responses
• Bilingual support: fluent in Thai and English
• Cultural fit: tuned for Thai context in daily-life and conversational settings

Benchmarks

We evaluated ChindaGo-4B using LLM-as-a-judge with openai/gpt-oss-120b as the evaluation model.
The benchmarks focused on daily life Q&A, emergency guidance, and cultural alignment in Thai language, comparing against its base model.

[Benchmark]

Training Data

The model was fine-tuned on a mixture of curated daily-life datasets and dedicated identification dataset was used to improve personalization and assistant-like behavior.
Datasets will be released later.

Example Usage (MLX)

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("yacht/ChindaGo-4B-MLX-4bit")

prompt = "ช่วยทำเช็กลิสต์ของใช้ฉุกเฉินในบ้านให้หน่อย"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)