---
license: mit
base_model: microsoft/MediPhi-Instruct
library_name: mlx
pipeline_tag: text-generation
language:
- en

tags:
- mlx
- mlx-lm
- phi
- phi-3
- medical
- clinical
- healthcare
- quantized
- 4bit
- on-device
- ios
- apple-silicon

model_type: phi
quantization:
- 4bit

datasets:
- microsoft/mediflow
- ncbi/pubmed
- epfl-llm/guidelines
- starmpcc/Asclepius-Synthetic-Clinical-Notes
- akemiH/NoteChat
- zhengyun21/PMC-Patients
- jpcorb20/medical_wikipedia
---

# MediPhi-Instruct (MLX · 4-bit)

This repository contains an **MLX-format 4-bit quantized** version of  
[`microsoft/MediPhi-Instruct`](https://huggingface.co/microsoft/MediPhi-Instruct),  
converted using **`mlx-lm`** for efficient **on-device inference** on Apple silicon.

This model is intended for **iOS / iPadOS / macOS** usage where memory and power
constraints require aggressive quantization while preserving clinical reasoning quality.

---

## Model details

- **Base model:** MediPhi-Instruct (Phi-3 family)
- **Parameters:** ~3.8B
- **Quantization:** 4-bit (MLX)
- **Format:** MLX (not GGUF)
- **Intended use:** On-device medical and clinical QA, decision support, and explanations
- **Language:** English

> ⚠️ This is a **conversion only**. No additional fine-tuning was performed.

---

## Why MLX 4-bit?

Compared to larger 4–7B medical models, MediPhi-Instruct shows:
- Strong clinical reasoning per parameter
- Better robustness under 4-bit quantization
- Lower memory footprint suitable for mobile devices

This makes it a strong candidate for **on-device medical assistants** on iPhone and iPad.

---

## Usage (MLX-LM)

### Install
```bash
pip install mlx-lm