diff --git a/bioimageio/engine/ray_apps/cellpose/__init__.py b/bioimageio/engine/ray_apps/cellpose/__init__.py
index a96ce04..6f1330d 100644
--- a/bioimageio/engine/ray_apps/cellpose/__init__.py
+++ b/bioimageio/engine/ray_apps/cellpose/__init__.py
@@ -1,17 +1,52 @@
 from hypha_rpc import api
+import numpy as np
 
 class CellposeModel:
     def __init__(self):
-        # Load model
-        import torch
+        from cellpose import core
+        # Check if GPU is available
+        self.use_GPU = core.use_gpu()
+        print('>>> GPU activated? %d' % self.use_GPU)
+        
+        # Initialize model caching attributes
+        self.cached_model_type = None
+        self.model = None
 
-    def predict(self, image: str) -> str:
-        prediction = "prediction of cellpose model on image: " + image
-        return prediction
+    def _load_model(self, model_type):
+        from cellpose import models
+        if self.model is None or model_type != self.cached_model_type:
+            print(f'Loading model: {model_type}')
+            self.model = models.Cellpose(gpu=self.use_GPU, model_type=model_type)
+            self.cached_model_type = model_type
+        else:
+            print(f'Reusing cached model: {model_type}')
+        return self.model
 
-    def train(self, data: str, config: str) -> str:
-        training = "training cellpose model on data: " + data + "with config:" + config
-        return training
+    def predict(self, images: list[np.ndarray], channels=None, diameter=None, flow_threshold=None, model_type='cyto3'):
+        """Run segmentation on the provided images using the specified model type."""
+        # Load the model, utilizing caching
+        model = self._load_model(model_type)
 
+        if channels is None:
+            # Default channels if not provided
+            channels = [[2, 3]] * len(images)
+        
+        # Perform segmentation using the model
+        masks, flows, styles, diams = model.eval(images, diameter=diameter, flow_threshold=flow_threshold, channels=channels)
+        
+        # Prepare the response with masks and diameters
+        results = {
+            'masks': [mask.tolist() for mask in masks],  # Converting numpy arrays to lists for JSON serialization
+            'diameters': diams  # List of estimated diameters for each image
+        }
 
+        return results
+
+    def train(self, images, labels, config):
+        """Train the model using the provided images and labels."""
+        # This method would handle the training process.
+        # Currently, returning a placeholder response.
+        raise NotImplementedError("Training functionality not implemented yet")
+
+# Export the CellposeModel class using Hypha RPC API
 api.export(CellposeModel)
diff --git a/bioimageio/engine/ray_apps/cellpose/manifest.yaml b/bioimageio/engine/ray_apps/cellpose/manifest.yaml
index c8769cf..87c43fc 100644
--- a/bioimageio/engine/ray_apps/cellpose/manifest.yaml
+++ b/bioimageio/engine/ray_apps/cellpose/manifest.yaml
@@ -8,6 +8,8 @@ ray_serve_config:
     num_gpus: 1
     runtime_env:
       pip:
+        - opencv-python-headless==4.2.0.34
+        - cellpose==3.0.11
         - torch==2.3.1
         - torchvision==0.18.1 
   autoscaling_config: