diff --git a/setup.sh b/setup.sh
index 943d3a4d..2f6a92ac 100755
--- a/setup.sh
+++ b/setup.sh
@@ -1,42 +1,101 @@
-#!/bin/bash
+
+#!/usr/bin/env bash
+
+# setup.sh
+# Cross-platform setup script for nejm-brain-to-text
+# This script creates a conda env and installs Python deps.
+# It detects macOS (Apple Silicon / Intel) and Linux and installs
+# an appropriate PyTorch build: CUDA wheel for Linux with CUDA,
+# and macOS (MPS) compatible wheel for Apple machines.
 
 # Ensure conda is available
+if ! command -v conda >/dev/null 2>&1; then
+    echo "conda not found in PATH. Please install Miniconda/Anaconda and retry."
+    exit 1
+fi
+
 source "$(conda info --base)/etc/profile.d/conda.sh"
 
-# Create conda environment with Python 3.10
-conda create -n b2txt25 python=3.10 -y
+# Create conda environment with Python 3.10 (idempotent)
+ENV_NAME=b2txt25
+PY_VER=3.10
+if conda env list | grep -q "^${ENV_NAME}[[:space:]]"; then
+    echo "Conda env '${ENV_NAME}' already exists. Skipping creation."
+else
+    conda create -n ${ENV_NAME} python=${PY_VER} -y
+fi
 
 # Activate the new environment
-conda activate b2txt25
+conda activate ${ENV_NAME}
 
 # Upgrade pip
-pip install --upgrade pip
-
-# Install PyTorch with CUDA 12.6
-pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
-
-# Install additional packages
-# TODO: remove redis
-pip install \
-    redis==5.2.1 \
-    jupyter==1.1.1 \
-    numpy==2.1.2 \
-    pandas==2.3.0 \
-    matplotlib==3.10.1 \
-    scipy==1.15.2 \
-    scikit-learn==1.6.1 \
-    tqdm==4.67.1 \
-    g2p_en==2.1.0 \
-    h5py==3.13.0 \
-    omegaconf==2.3.0 \
-    editdistance==0.8.1 \
-    -e . \
-    huggingface-hub==0.33.1 \
-    transformers==4.53.0 \
-    tokenizers==0.21.2 \
-    accelerate==1.8.1 \
-    bitsandbytes==0.46.0
+python -m pip install --upgrade pip
+
+# Detect platform and choose PyTorch install
+OS_NAME=$(uname -s)
+ARCH=$(uname -m)
+
+echo "Detected OS=${OS_NAME} ARCH=${ARCH}"
+
+if [[ "${OS_NAME}" == "Darwin" ]]; then
+    # macOS: prefer PyTorch with MPS (Apple Metal Performance Shaders)
+    # Use the stable cpu/macos build from PyPI which enables MPS on Apple Silicon
+    echo "Installing PyTorch (macOS/MPS-compatible)..."
+    # Recent PyTorch stable releases provide macOS wheels on PyPI.
+    python -m pip install --upgrade "torch" torchvision torchaudio
+    EXTRA_PIP_PACKAGES=(
+        redis==5.2.1
+        jupyter==1.1.1
+        numpy==2.1.2
+        pandas==2.3.0
+        matplotlib==3.10.1
+        scipy==1.15.2
+        scikit-learn==1.6.1
+        tqdm==4.67.1
+        g2p_en==2.1.0
+        h5py==3.13.0
+        omegaconf==2.3.0
+        editdistance==0.8.1
+        -e .
+        huggingface-hub==0.33.1
+        transformers==4.53.0
+        tokenizers==0.21.2
+        accelerate==1.8.1
+    )
+    # bitsandbytes is not supported on macOS in many setups; leave it out by default
+    echo "Note: 'bitsandbytes' is skipped on macOS by default as it's typically Linux/CUDA-only."
+
+else
+    # Assume Linux by default
+    echo "Installing PyTorch (Linux/CUDA) using index-url for CUDA 12.6..."
+    python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
+    EXTRA_PIP_PACKAGES=(
+        redis==5.2.1
+        jupyter==1.1.1
+        numpy==2.1.2
+        pandas==2.3.0
+        matplotlib==3.10.1
+        scipy==1.15.2
+        scikit-learn==1.6.1
+        tqdm==4.67.1
+        g2p_en==2.1.0
+        h5py==3.13.0
+        omegaconf==2.3.0
+        editdistance==0.8.1
+        -e .
+        huggingface-hub==0.33.1
+        transformers==4.53.0
+        tokenizers==0.21.2
+        accelerate==1.8.1
+        bitsandbytes==0.46.0
+    )
+fi
 
 echo
-echo "Setup complete! Verify it worked by activating the conda environment with the command 'conda activate b2txt25'."
+echo "Installing additional Python packages..."
+python -m pip install "${EXTRA_PIP_PACKAGES[@]}"
+
 echo
+echo "Setup complete!"
+echo "Verify it worked by activating the conda environment with: conda activate ${ENV_NAME}"
+echo "On macOS, to use MPS (Metal) backend in PyTorch, set device='mps' in your code or use torch.device('mps') when available."