diff --git a/machine-learning/export/env.yaml b/machine-learning/export/env.yaml
index f53f5b001..c1f80b4b0 100644
--- a/machine-learning/export/env.yaml
+++ b/machine-learning/export/env.yaml
@@ -22,5 +22,5 @@ dependencies:
   - pip:
       - multilingual-clip
       - onnx-simplifier
-      - tensorflow
+      - tensorflow==2.14.*
 category: main
diff --git a/machine-learning/export/models/tfclip.py b/machine-learning/export/models/tfclip.py
index 4dbe00d10..8dee62f4d 100644
--- a/machine-learning/export/models/tfclip.py
+++ b/machine-learning/export/models/tfclip.py
@@ -10,7 +10,7 @@ from .util import ModelType, get_model_path
 class _CLIPWrapper(tf.Module):
     def __init__(self, model_name: str):
         super(_CLIPWrapper)
-        self.model = TFCLIPModel.from_pretrained(model_name)
+        self.model = TFCLIPModel.from_pretrained(model_name).half()
 
     @tf.function()
     def encode_image(self, input):
@@ -22,11 +22,13 @@ class _CLIPWrapper(tf.Module):
 
 
 # exported model signatures use batch size 2 because of the following reasons:
-# 1. ARM-NN cannot use dynamic batch sizes
+# 1. ARM-NN cannot use dynamic batch sizes for complex models like CLIP ViT
 # 2. batch size 1 creates a larger TF-Lite model that uses a lot (50%) more RAM
-# 3. batch size 2 is ~50% faster on GPU than 1 while 4 (or larger) are not faster
+# 3. batch size 2 is ~50% faster on GPU than 1 while 4 (or larger) are not really faster
 # 4. batch size >2 wastes more computation if only a single image is processed
-BATCH_SIZE = 2
+BATCH_SIZE_IMAGE = 2
+# On most small-scale systems there will only be one query at a time, no sense in batching
+BATCH_SIZE_TEXT = 1
 
 SIGNATURE_TEXT = "encode_text"
 SIGNATURE_IMAGE = "encode_image"
@@ -52,12 +54,12 @@ def _export_temporary_tf_model(model_name, tmp_path: str, context_length: int):
     wrapper = _CLIPWrapper(model_name)
     conf = wrapper.model.config.vision_config
     spec_visual = tf.TensorSpec(
-        shape=(BATCH_SIZE, conf.num_channels, conf.image_size, conf.image_size), dtype=tf.float32
+        shape=(BATCH_SIZE_IMAGE, conf.num_channels, conf.image_size, conf.image_size), dtype=tf.float32
     )
     encode_image = wrapper.encode_image.get_concrete_function(spec_visual)
-    spec_text = tf.TensorSpec(shape=(BATCH_SIZE, context_length), dtype=tf.int32)
+    spec_text = tf.TensorSpec(shape=(BATCH_SIZE_TEXT, context_length), dtype=tf.int32)
     encode_text = wrapper.encode_text.get_concrete_function(spec_text)
-    signatures = {"encode_text": encode_text, "encode_image": encode_image}
+    signatures = { SIGNATURE_IMAGE: encode_image, SIGNATURE_TEXT: encode_text}
     tf.saved_model.save(wrapper, tmp_path, signatures)
 
 
diff --git a/machine-learning/export/run.py b/machine-learning/export/run.py
index 49dfef5a1..6fa262ad0 100644
--- a/machine-learning/export/run.py
+++ b/machine-learning/export/run.py
@@ -4,9 +4,10 @@ from pathlib import Path
 from tempfile import TemporaryDirectory
 
 from huggingface_hub import create_repo, login, upload_folder
-from models import mclip, openclip, tfclip
 from rich.progress import Progress
 
+from models import mclip, openclip, tfclip
+
 models = [
     "RN50::openai",
     "RN50::yfcc15m",
diff --git a/machine-learning/poetry.lock b/machine-learning/poetry.lock
index 2b5a11e8b..d0c995952 100644
--- a/machine-learning/poetry.lock
+++ b/machine-learning/poetry.lock
@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 1.6.1 and should not be changed by hand.
+# This file is automatically @generated by Poetry 1.7.1 and should not be changed by hand.
 
 [[package]]
 name = "aiocache"
@@ -3882,6 +3882,30 @@ files = [
 [package.dependencies]
 mpmath = ">=0.19"
 
+[[package]]
+name = "tflite-runtime"
+version = "2.14.0"
+description = "TensorFlow Lite is for mobile and embedded devices."
+optional = false
+python-versions = "*"
+files = [
+    {file = "tflite_runtime-2.14.0-cp310-cp310-manylinux2014_x86_64.whl", hash = "sha256:bb11df4283e281cd609c621ac9470ad0cb5674408593272d7593a2c6bde8a808"},
+    {file = "tflite_runtime-2.14.0-cp310-cp310-manylinux_2_34_aarch64.whl", hash = "sha256:d38c6885f5e9673c11a61ccec5cad7c032ab97340718d26b17794137f398b780"},
+    {file = "tflite_runtime-2.14.0-cp310-cp310-manylinux_2_34_armv7l.whl", hash = "sha256:7fe33f763263d1ff2733a09945a7547ab063d8bc311fd2a1be8144d850016ad3"},
+    {file = "tflite_runtime-2.14.0-cp311-cp311-manylinux2014_x86_64.whl", hash = "sha256:195ab752e7e57329a68e54dd3dd5439fad888b9bff1be0f0dc042a3237a90e4d"},
+    {file = "tflite_runtime-2.14.0-cp311-cp311-manylinux_2_34_aarch64.whl", hash = "sha256:ce9fa5d770a9725c746dcbf6f59f3178233b3759f09982e8b2db8d2234c333b0"},
+    {file = "tflite_runtime-2.14.0-cp311-cp311-manylinux_2_34_armv7l.whl", hash = "sha256:c4e66a74165b18089c86788400af19fa551768ac782d231a9beae2f6434f7949"},
+    {file = "tflite_runtime-2.14.0-cp38-cp38-manylinux2014_x86_64.whl", hash = "sha256:9f965054467f7890e678943858c6ac76a5197b17f61b48dcbaaba0af41d541a7"},
+    {file = "tflite_runtime-2.14.0-cp38-cp38-manylinux_2_34_aarch64.whl", hash = "sha256:437167fe3d8b12f50f5d694da8f45d268ab84a495e24c3dd810e02e1012125de"},
+    {file = "tflite_runtime-2.14.0-cp38-cp38-manylinux_2_34_armv7l.whl", hash = "sha256:79d8e17f68cc940df7e68a177b22dda60fcffba195fb9dd908d03724d65fd118"},
+    {file = "tflite_runtime-2.14.0-cp39-cp39-manylinux2014_x86_64.whl", hash = "sha256:4aa740210a0fd9e4db4a46e9778914846b136e161525681b41575ca4896158fb"},
+    {file = "tflite_runtime-2.14.0-cp39-cp39-manylinux_2_34_aarch64.whl", hash = "sha256:be198b7dc4401204be54a15884d9e336389790eb707439524540f5a9329fdd02"},
+    {file = "tflite_runtime-2.14.0-cp39-cp39-manylinux_2_34_armv7l.whl", hash = "sha256:eca7672adca32727bbf5c0f1caf398fc17bbe222f2a684c7a2caea6fc6767203"},
+]
+
+[package.dependencies]
+numpy = ">=1.23.2"
+
 [[package]]
 name = "threadpoolctl"
 version = "3.2.0"
@@ -4025,6 +4049,14 @@ dev = ["tokenizers[testing]"]
 docs = ["setuptools_rust", "sphinx", "sphinx_rtd_theme"]
 testing = ["black (==22.3)", "datasets", "numpy", "pytest", "requests"]
 
+[[package]]
+name = "torch"
+version = "2.0.1"
+description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration"
+optional = false
+python-versions = "*"
+files = []
+
 [[package]]
 name = "torch"
 version = "2.1.0"
@@ -4772,4 +4804,4 @@ testing = ["coverage (>=5.0.3)", "zope.event", "zope.testing"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.11"
-content-hash = "bba5f87aa67bc1d2283a9f4b471ef78e572337f22413870d324e908014410d53"
+content-hash = "56614afdeeeec3b7f0b786771a8fcc126761c882b1033664056042833767e521"
diff --git a/machine-learning/pyproject.toml b/machine-learning/pyproject.toml
index cd4acf9be..32f84997f 100644
--- a/machine-learning/pyproject.toml
+++ b/machine-learning/pyproject.toml
@@ -29,6 +29,7 @@ python-multipart = "^0.0.6"
 orjson = "^3.9.5"
 safetensors = "0.3.2"
 gunicorn = "^21.1.0"
+tflite-runtime = "^2.14.0"
 
 [tool.poetry.group.dev.dependencies]
 mypy = "^1.3.0"