diff --git a/cht-llama-cpp/build-aarch64-apple-darwin.sh b/cht-llama-cpp/build-aarch64-apple-darwin.sh
index 121c777..7b226a9 100755
--- a/cht-llama-cpp/build-aarch64-apple-darwin.sh
+++ b/cht-llama-cpp/build-aarch64-apple-darwin.sh
@@ -1,16 +1,17 @@
 #!/usr/bin/env bash
 set -e
-export VERSION=1.1.2
+export VERSION=1.2.0
 
 test -f venv/bin/activate || python -m venv venv
 source venv/bin/activate
 
-pip install -r requirements.txt pyinstaller
+CMAKE_FORCE=1 CMAKE_ARGS="-DLLAMA_METAL=on" pip install -r requirements.txt pyinstaller
 LLAMA_CPP_PATH=$(python -c 'import llama_cpp; print(llama_cpp.__path__[0])')
 NAME=cht-llama-cpp-mistral-${VERSION}-aarch64-apple-darwin
 pyinstaller --onefile \
   --target-arch arm64 \
   --add-binary "$LLAMA_CPP_PATH/libllama.dylib:llama_cpp" \
+  --add-binary "$LLAMA_CPP_PATH/ggml-metal.metal:llama_cpp" \
   --name=$NAME \
   main.py
 cp dist/$NAME dist/cht-llama-cpp-mistral-${VERSION%%.*}-aarch64-apple-darwin
diff --git a/cht-llama-cpp/build.sh b/cht-llama-cpp/build.sh
index 8dd8420..6607fc1 100755
--- a/cht-llama-cpp/build.sh
+++ b/cht-llama-cpp/build.sh
@@ -1,7 +1,8 @@
 #!/bin/bash
 set -e
-export VERSION=1.1.2
+export VERSION=1.2.0
 source "$(dirname "${BASH_SOURCE[0]}")/../utils.sh"
 
 build_cpu ghcr.io/premai-io/chat-mistral-7b-instruct-q5 mistral-7b-instruct-v0.1.Q5_0 --build-arg="MODEL_ID=mistral-7b-instruct-v0.1.Q5_0" --build-arg="MODEL_DOWNLOAD_URL=https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.1-GGUF/resolve/main/mistral-7b-instruct-v0.1.Q5_0.gguf" ${@:1}
 build_cpu ghcr.io/premai-io/chat-mistral-7b-openorca-q5 mistral-7b-openorca.Q5_K_S    --build-arg="MODEL_ID=mistral-7b-openorca.Q5_K_S"    --build-arg="MODEL_DOWNLOAD_URL=https://huggingface.co/TheBloke/Mistral-7B-OpenOrca-GGUF/resolve/main/mistral-7b-openorca.Q5_K_S.gguf"         ${@:1}
+build_cpu ghcr.io/premai-io/chat-mistral-7b-yarn-q4 yarn-mistral-7b-128k.Q4_K_M    --build-arg="MODEL_ID=yarn-mistral-7b-128k.Q4_K_M"    --build-arg="MODEL_DOWNLOAD_URL=https://huggingface.co/TheBloke/Yarn-Mistral-7B-128k-GGUF/blob/main/yarn-mistral-7b-128k.Q3_K_M.gguf"         ${@:1}
diff --git a/cht-llama-cpp/docker/cpu/Dockerfile b/cht-llama-cpp/docker/cpu/Dockerfile
index 8344fde..74d0385 100644
--- a/cht-llama-cpp/docker/cpu/Dockerfile
+++ b/cht-llama-cpp/docker/cpu/Dockerfile
@@ -15,4 +15,5 @@ RUN wget -O ./ml/models/${MODEL_ID}.gguf ${MODEL_DOWNLOAD_URL}
 
 COPY . .
 ENV MODEL_ID=$MODEL_ID
+ENV DEVICE=cpu
 CMD python main.py
diff --git a/cht-llama-cpp/main.py b/cht-llama-cpp/main.py
index 6a5d41b..72ababe 100644
--- a/cht-llama-cpp/main.py
+++ b/cht-llama-cpp/main.py
@@ -13,17 +13,18 @@
 MODEL_PATH = f"./ml/models/{os.getenv('MODEL_ID', 'yarn-mistral-7b-128k.Q4_K_M')}.gguf"
 # Mistral gguf follows ChatML syntax
 # https://github.com/openai/openai-python/blob/main/chatml.md
-
+PROMPT_TEMPLATE_STRING = '{"system_prompt_template": "<|im_start|>system\\n{}\\n<|im_end|>\\n", "default_system_text": "You are an helpful AI assistant.", "user_prompt_template": "<|im_start|>user\\n{}\\n<|im_end|>\\n", "assistant_prompt_template": "<|im_start|>assistant\\n{}\\n<|im_end|>\\n", "request_assistant_response_token": "<|im_start|>assistant\\n", "template_format": "chatml"}'  # noqa
+MODEL_CTX = 4096
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--model-path", help="Path to GGUF", default=MODEL_PATH)
     parser.add_argument("--port", help="Port to run model server on", type=int, default=8000)
-    parser.add_argument("--ctx", help="Context dimension", type=int, default=4096)
+    parser.add_argument("--ctx", help="Context dimension", type=int, default=MODEL_CTX)
     parser.add_argument(
         "--prompt_template",
         help="Prompt Template",
         type=str,
-        default='{"system_prompt_template": "<|im_start|>system\\n{}\\n<|im_end|>\\n", "default_system_text": "You are an helpful AI assistant.", "user_prompt_template": "<|im_start|>user\\n{}\\n<|im_end|>\\n", "assistant_prompt_template": "<|im_start|>assistant\\n{}\\n<|im_end|>\\n", "request_assistant_response_token": "<|im_start|>assistant\\n", "template_format": "chatml"}',  # noqa
+        default=PROMPT_TEMPLATE_STRING,  # noqa
     )  # noqa
     args = parser.parse_args()
     MODEL_PATH = args.model_path
diff --git a/cht-llama-cpp/models.py b/cht-llama-cpp/models.py
index af4b944..6f34a64 100644
--- a/cht-llama-cpp/models.py
+++ b/cht-llama-cpp/models.py
@@ -1,5 +1,6 @@
 import json
 import multiprocessing
+import os
 from typing import Any, Dict, List
 
 from llama_cpp import Llama, llama_chat_format, llama_types
@@ -66,7 +67,8 @@ def get_model(cls, model_path, prompt_template_jsonstr, n_ctx):
             cls.PROMPT_TEMPLATE = json.loads(prompt_template_jsonstr)
             chat_format = cls.PROMPT_TEMPLATE.get("template_format", "chatml")
         if cls.model is None:
-            cls.model = Llama(model_path, chat_format=chat_format, n_ctx=n_ctx)
+            gpu_offload_layers = -1 if os.getenv("DEVICE") != "cpu" else 0
+            cls.model = Llama(model_path, chat_format=chat_format, n_ctx=n_ctx, n_gpu_layers=gpu_offload_layers)
 
         return cls.model
 
diff --git a/cht-llama-cpp/requirements.txt b/cht-llama-cpp/requirements.txt
index cf7f103..38f208f 100644
--- a/cht-llama-cpp/requirements.txt
+++ b/cht-llama-cpp/requirements.txt
@@ -6,4 +6,4 @@ tqdm==4.65.0
 httpx==0.23.3
 python-dotenv==1.0.0
 tenacity==8.2.2
-llama-cpp-python==0.2.14
+llama-cpp-python==0.2.18