---
language: en
tags:
- mlx
- qwen3.5
- nutrition
- food.com
pipeline_tag: image-text-to-text
library_name: mlx
license: apache-2.0
base_model:
- Qwen/Qwen3.5-2B
---

# Qwen3.5 Nutrition Finte Tune

This is a highly specialized, local-first Vision-Language Model (VLM) designed to estimate nutritional macros and physical weight from images of food. The model is fine-tuned to output strict, minified JSON payloads containing calories, macros (protein, fat, carbohydrates), and estimated dry/physical weight in grams. 

It has been natively converted to Apple's MLX format and quantized to 4-bit (INT4), reducing its memory footprint to roughly ~1.5GB. This allows for lightning-fast, entirely offline, on-device inference on Apple Silicon hardware (macOS, iOS, iPadOS) using the Neural Engine.

- **Base Architecture:** [Qwen3.5-2B](https://huggingface.co/Qwen) (Omni/Unified Vision-Language architecture)
- **Framework:** MLX (`mlx-vlm`)
- **Quantization:** INT4 (4-bit)
- **Fine-Tuning Method:** LoRA (via Unsloth), merged to 16-bit prior to MLX conversion.
- **Primary Task:** Multimodal Image-to-JSON parsing.

## Intended Use
- **Primary Use Case:** Seamless integration into native Swift/Apple ecosystem applications requiring on-device visual food analysis without relying on cloud APIs.
- **Input:** A user prompt (e.g. food description) + a standard RGB food image.
- **Output:** A standardized JSON schema containing absolute nutritional values.

## Training Data & Credit
This model was fine-tuned on a custom dataset synthesized from a massive corpus of recipe data and user-uploaded images.
- **Dataset Source:** Modified from the [Food.com Recipes and Reviews](https://www.kaggle.com/datasets/irkaal/foodcom-recipes-and-reviews) dataset. The text macros were mathematically processed to derive estimated physical weights, combined with dynamically downloaded web images, and formatted into an OpenAI-style JSONL conversation structure.
- **Dataset License:** Public Domain (CC0)

## Licensing & Attribution
- **Model License:** [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0) (Inherited from the Qwen 3.5 2B base model).
- **Attribution:** All base architectural credit belongs to the Qwen team at Alibaba Cloud. 

## Usage Example (MLX)
```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import get_message_profile
from mlx_vlm.utils import load_image

# Load the INT4 MLX model
model, processor = load("alexsofonea/Qwen3.5_2B_MLX_Nutrition")

# Format input
image = load_image("path/to/food.jpg")
prompt = "Output JSON.\nEstimate."

# Generate JSON
output = generate(model, processor, prompt, image, max_tokens=150, temperature=0.0)
print(output)
```

## About Developer
```JSON
{
  "developer": {
    "name": "Alex",
    "age": 19,
    "location": "Munchen, Germany",
    "company": "Tecky",
    "role": "Founder",
    "background": [
      "Developer since age 7",
      "Filmmaker"
    ]
  },
  "mission": "Building local-first AI agents and integrating modern cinematic design into native software."
}
```
Contact:
- [alex@alexsofonea.com](mailto:alex@alexsofonea.com)
- [LinkedIn](https://www.linkedin.com/in/alexsofonea/)
- [Website](https://alexsofonea.com)