diff --git a/download_and_run.sh b/download_and_run.sh
new file mode 100755
index 0000000..867cb87
--- /dev/null
+++ b/download_and_run.sh
@@ -0,0 +1,160 @@
+#!/bin/bash
+set -e
+
+# Default values
+# MIRROR_URL="https://hf-mirror.com"
+DEFAULT_SAVE_DIR="$HOME/.aha"
+MODEL_ALIAS=""
+
+# Help function
+show_help() {
+    echo "Usage: $0 [model_alias]"
+    echo ""
+    echo "Arguments:"
+    echo "  model_alias       The model alias to download (e.g., voxcpm, qwen2.5vl-3b)"
+    echo ""
+    echo "Available models:"
+    echo "  minicpm4-0.5b"
+    echo "  qwen2.5vl-3b"
+    echo "  qwen2.5vl-7b"
+    echo "  qwen3vl-2b"
+    echo "  qwen3vl-4b"
+    echo "  qwen3vl-8b"
+    echo "  qwen3vl-32b"
+    echo "  deepseek-ocr"
+    echo "  hunyuan-ocr"
+    echo "  paddleocr-vl"
+    echo "  RMBG2.0"
+    echo "  voxcpm"
+    echo "  voxcpm1.5"
+    echo "  glm-asr-nano-2512"
+    echo ""
+    exit 1
+}
+
+# Check if model alias is provided
+if [ -z "$1" ]; then
+    show_help
+fi
+
+MODEL_ALIAS=$1
+
+# Map alias to Repo ID
+MODEL_ID=""
+case $MODEL_ALIAS in
+    "minicpm4-0.5b")
+        MODEL_ID="OpenBMB/MiniCPM4-0.5B"
+        ;;
+    "qwen2.5vl-3b")
+        MODEL_ID="Qwen/Qwen2.5-VL-3B-Instruct"
+        ;;
+    "qwen2.5vl-7b")
+        MODEL_ID="Qwen/Qwen2.5-VL-7B-Instruct"
+        ;;
+    "qwen3vl-2b")
+        MODEL_ID="Qwen/Qwen3-VL-2B-Instruct"
+        ;;
+    "qwen3vl-4b")
+        MODEL_ID="Qwen/Qwen3-VL-4B-Instruct"
+        ;;
+    "qwen3vl-8b")
+        MODEL_ID="Qwen/Qwen3-VL-8B-Instruct"
+        ;;
+    "qwen3vl-32b")
+        MODEL_ID="Qwen/Qwen3-VL-32B-Instruct"
+        ;;
+    "deepseek-ocr")
+        MODEL_ID="deepseek-ai/DeepSeek-OCR"
+        ;;
+    "hunyuan-ocr")
+        MODEL_ID="Tencent-Hunyuan/HunyuanOCR"
+        ;;
+    "paddleocr-vl")
+        MODEL_ID="PaddlePaddle/PaddleOCR-VL"
+        ;;
+    "RMBG2.0")
+        MODEL_ID="AI-ModelScope/RMBG-2.0"
+        ;;
+    "voxcpm")
+        MODEL_ID="OpenBMB/VoxCPM-0.5B"
+        ;;
+    "voxcpm1.5")
+        MODEL_ID="OpenBMB/VoxCPM1.5"
+        ;;
+    "glm-asr-nano-2512")
+        MODEL_ID="zai-org/GLM-ASR-Nano-2512"
+        ;;
+    *)
+        echo "Error: Unknown model alias '$MODEL_ALIAS'"
+        show_help
+        ;;
+esac
+
+echo "Selected Model: $MODEL_ALIAS"
+echo "Repo ID: $MODEL_ID"
+echo "Target Directory: $DEFAULT_SAVE_DIR/$MODEL_ID"
+
+# Prepare environment for acceleration (Default to mirror if not set)
+if [ -z "$HF_ENDPOINT" ]; then
+    export HF_ENDPOINT=$MIRROR_URL
+    echo "Using default HF Mirror: $HF_ENDPOINT"
+else
+    echo "Using custom HF Endpoint: $HF_ENDPOINT"
+fi
+
+# Check if huggingface-cli and hf_transfer are installed
+CLI_CMD=""
+if command -v huggingface-cli &> /dev/null; then
+    CLI_CMD="huggingface-cli"
+elif command -v hf &> /dev/null; then
+    CLI_CMD="hf"
+else
+    echo "huggingface-cli not found. Installing via pip..."
+    if command -v pip3 &> /dev/null; then
+        pip3 install -U "huggingface_hub[cli]" hf_transfer
+    elif command -v pip &> /dev/null; then
+        pip install -U "huggingface_hub[cli]" hf_transfer
+    else
+        echo "Error: pip is not available. Please install python and pip."
+        exit 1
+    fi
+    CLI_CMD="huggingface-cli"
+fi
+
+# Try to install hf_transfer if simple python check fails (optional but recommended for speed)
+if ! python3 -c "import hf_transfer" &> /dev/null; then
+     echo "Installing hf_transfer for faster downloads..."
+     pip3 install -U hf_transfer || echo "Warning: hf_transfer install failed, falling back to standard download."
+fi
+
+# Enable HF Transfer (Rust-based downloader) - Default to 1 (On) unless explicitly disabled
+export HF_HUB_ENABLE_HF_TRANSFER=${HF_HUB_ENABLE_HF_TRANSFER:-1}
+if [ "$HF_HUB_ENABLE_HF_TRANSFER" == "1" ]; then
+    echo "Enabled HF_HUB_ENABLE_HF_TRANSFER for high-speed download"
+else
+    echo "HF_HUB_ENABLE_HF_TRANSFER disabled. Using standard Python downloader."
+fi
+
+# Download model
+echo "Downloading model with acceleration via $CLI_CMD..."
+# Note: --resume-download is deprecated/implicit in newer versions, removing it. 
+# --local-dir-use-symlinks matches standard CLI if version is recent. 
+if ! $CLI_CMD download "$MODEL_ID" --local-dir "$DEFAULT_SAVE_DIR/$MODEL_ID" --local-dir-use-symlinks False $TOKEN_ARG; then
+    echo ""
+    echo "Error: Download failed."
+    echo "If you received a 429 Rate Limit error, please provide a Hugging Face Token."
+    echo "Usage: ./download_and_run.sh $MODEL_ALIAS [hf_token]"
+    echo "Or set the HF_TOKEN environment variable."
+    exit 1
+fi
+
+# Run cargo
+echo "Starting application with cargo..."
+# Detect if on Mac to add 'metal' feature
+if [[ "$OSTYPE" == "darwin"* ]]; then
+    echo "Detected MacOS. Running with 'metal' feature..."
+    cargo run -r --features metal -- --model "$MODEL_ALIAS" --weight-path "$DEFAULT_SAVE_DIR/$MODEL_ID"
+else
+    echo "Running with default features (cuda)..."
+    cargo run -r --features cuda -- --model "$MODEL_ALIAS" --weight-path "$DEFAULT_SAVE_DIR/$MODEL_ID"
+fi
diff --git a/src/models/voxcpm/generate.rs b/src/models/voxcpm/generate.rs
index a7dee6c..0071a81 100644
--- a/src/models/voxcpm/generate.rs
+++ b/src/models/voxcpm/generate.rs
@@ -199,6 +199,10 @@ impl VoxCPMGenerate {
         )?;
         Ok(audio)
     }
+
+    pub fn sample_rate(&self) -> usize {
+        self.sample_rate
+    }
 }
 
 impl GenerateModel for VoxCPMGenerate {
diff --git a/tests/test_deepseek_ocr.rs b/tests/test_deepseek_ocr.rs
index e73dee1..a7b0a24 100644
--- a/tests/test_deepseek_ocr.rs
+++ b/tests/test_deepseek_ocr.rs
@@ -32,10 +32,12 @@ fn deepseek_ocr_generate() -> Result<()> {
         "metadata": {"base_size": "640", "image_size": "640", "crop_mode": "false"}
     }
     "#;
-    let model_path = "/home/jhq/huggingface_model/deepseek-ai/DeepSeek-OCR/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/deepseek-ai/DeepSeek-OCR/", save_dir);
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = DeepseekOCRGenerateModel::init(model_path, None, None)?;
+    let mut model = DeepseekOCRGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let i_start = Instant::now();
@@ -84,10 +86,12 @@ async fn deepseek_ocr_stream() -> Result<()> {
         "metadata": {"base_size": "640", "image_size": "640", "crop_mode": "false"}
     }
     "#;
-    let model_path = "/home/jhq/huggingface_model/deepseek-ai/DeepSeek-OCR/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/deepseek-ai/DeepSeek-OCR/", save_dir);
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = DeepseekOCRGenerateModel::init(model_path, None, None)?;
+    let mut model = DeepseekOCRGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let mut stream = pin!(model.generate_stream(mes)?);
diff --git a/tests/test_gelab_zero.rs b/tests/test_gelab_zero.rs
index 4aff3e8..211f4e5 100644
--- a/tests/test_gelab_zero.rs
+++ b/tests/test_gelab_zero.rs
@@ -8,7 +8,9 @@ use anyhow::Result;
 fn gelab_zero_generate() -> Result<()> {
     // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda gelab_zero_generate -r -- --nocapture
 
-    let model_path = "/home/jhq/huggingface_model/stepfun-ai/GELab-Zero-4B-preview";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/stepfun-ai/GELab-Zero-4B-preview", save_dir);
 
     let message = r#"
     {
@@ -56,7 +58,7 @@ fn gelab_zero_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut qwen3vl = Qwen3VLGenerateModel::init(model_path, None, None)?;
+    let mut qwen3vl = Qwen3VLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
diff --git a/tests/test_glm_asr_nano.rs b/tests/test_glm_asr_nano.rs
index 7414cb5..b0cf26b 100644
--- a/tests/test_glm_asr_nano.rs
+++ b/tests/test_glm_asr_nano.rs
@@ -8,7 +8,9 @@ use rocket::futures::StreamExt;
 #[test]
 fn glm_asr_nano_generate() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda glm_asr_nano_generate -r -- --nocapture
-    let model_path = "/home/jhq/huggingface_model/ZhipuAI/GLM-ASR-Nano-2512/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/ZhipuAI/GLM-ASR-Nano-2512/", save_dir);
     let message = r#"
     {
         "model": "glm-asr-nano",
@@ -34,7 +36,7 @@ fn glm_asr_nano_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut glm_asr_model = GlmAsrNanoGenerateModel::init(model_path, None, None)?;
+    let mut glm_asr_model = GlmAsrNanoGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let i_start = Instant::now();
@@ -54,7 +56,9 @@ fn glm_asr_nano_generate() -> Result<()> {
 #[tokio::test]
 async fn glm_asr_nano_stream() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda glm_asr_nano_stream -r -- --nocapture
-    let model_path = "/home/jhq/huggingface_model/ZhipuAI/GLM-ASR-Nano-2512/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/ZhipuAI/GLM-ASR-Nano-2512/", save_dir);
     let message = r#"
     {
         "model": "glm-asr-nano",
@@ -80,7 +84,7 @@ async fn glm_asr_nano_stream() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut glm_asr_model = GlmAsrNanoGenerateModel::init(model_path, None, None)?;
+    let mut glm_asr_model = GlmAsrNanoGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let i_start = Instant::now();
diff --git a/tests/test_hunyuan_ocr.rs b/tests/test_hunyuan_ocr.rs
index adf887e..fa42c4d 100644
--- a/tests/test_hunyuan_ocr.rs
+++ b/tests/test_hunyuan_ocr.rs
@@ -31,10 +31,12 @@ fn hunyuan_ocr_generate() -> Result<()> {
         ]
     }
     "#;
-    let model_path = "/home/jhq/huggingface_model/Tencent-Hunyuan/HunyuanOCR/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Tencent-Hunyuan/HunyuanOCR/", save_dir);
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = HunyuanOCRGenerateModel::init(model_path, None, None)?;
+    let mut model = HunyuanOCRGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let i_start = Instant::now();
@@ -79,10 +81,12 @@ async fn hunyuan_ocr_stream() -> Result<()> {
         ]
     }
     "#;
-    let model_path = "/home/jhq/huggingface_model/Tencent-Hunyuan/HunyuanOCR/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Tencent-Hunyuan/HunyuanOCR/", save_dir);
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = HunyuanOCRGenerateModel::init(model_path, None, None)?;
+    let mut model = HunyuanOCRGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let mut stream = pin!(model.generate_stream(mes)?);
diff --git a/tests/test_minicpm4.rs b/tests/test_minicpm4.rs
index 5477e17..72636b3 100644
--- a/tests/test_minicpm4.rs
+++ b/tests/test_minicpm4.rs
@@ -11,7 +11,9 @@ fn minicpm_generate() -> Result<()> {
     // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda minicpm_generate -r -- --nocapture
     // test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn minicpm_generate -r -- --nocapture
 
-    let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/MiniCPM4-0.5B/", save_dir);
     let message = r#"
     {
         "temperature": 0.3,
@@ -27,7 +29,7 @@ fn minicpm_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = MiniCPMGenerateModel::init(model_path, None, None)?;
+    let mut model = MiniCPMGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
@@ -50,7 +52,9 @@ fn minicpm_generate() -> Result<()> {
 async fn minicpm_stream() -> Result<()> {
     // test with cuda+flash-attn: RUST_BACKTRACE=1 cargo test -F cuda,flash-attn minicpm_stream -r -- --nocapture
 
-    let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/MiniCPM4-0.5B/", save_dir);
 
     let message = r#"
     {
@@ -65,7 +69,7 @@ async fn minicpm_stream() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = MiniCPMGenerateModel::init(model_path, None, None)?;
+    let mut model = MiniCPMGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
diff --git a/tests/test_paddleocr_vl.rs b/tests/test_paddleocr_vl.rs
index 834a396..682afb3 100644
--- a/tests/test_paddleocr_vl.rs
+++ b/tests/test_paddleocr_vl.rs
@@ -32,10 +32,12 @@ fn paddleocr_vl_generate() -> Result<()> {
         "stream": false
     }
     "#;
-    let model_path = "/home/jhq/huggingface_model/PaddlePaddle/PaddleOCR-VL/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/PaddlePaddle/PaddleOCR-VL/", save_dir);
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = PaddleOCRVLGenerateModel::init(model_path, None, None)?;
+    let mut model = PaddleOCRVLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let i_start = Instant::now();
@@ -79,10 +81,12 @@ async fn paddleocr_vl_stream() -> Result<()> {
         ]
     }
     "#;
-    let model_path = "/home/jhq/huggingface_model/PaddlePaddle/PaddleOCR-VL/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/PaddlePaddle/PaddleOCR-VL/", save_dir);
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = PaddleOCRVLGenerateModel::init(model_path, None, None)?;
+    let mut model = PaddleOCRVLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
     let mut stream = pin!(model.generate_stream(mes)?);
diff --git a/tests/test_qwen2_5vl.rs b/tests/test_qwen2_5vl.rs
index d0423be..886f03c 100644
--- a/tests/test_qwen2_5vl.rs
+++ b/tests/test_qwen2_5vl.rs
@@ -13,7 +13,9 @@ fn qwen2_5vl_generate() -> Result<()> {
     // let device = Device::cuda_if_available(0)?;
     // let dtype = DType::BF16;
 
-    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen2.5-VL-3B-Instruct/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Qwen/Qwen2.5-VL-3B-Instruct/", save_dir);
 
     let message = r#"
     {
@@ -40,7 +42,7 @@ fn qwen2_5vl_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = Qwen2_5VLGenerateModel::init(model_path, None, None)?;
+    let mut model = Qwen2_5VLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
@@ -65,7 +67,9 @@ async fn qwen2_5vl_stream() -> Result<()> {
     // let device = Device::cuda_if_available(0)?;
     // let dtype = DType::BF16;
 
-    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen2.5-VL-3B-Instruct/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Qwen/Qwen2.5-VL-3B-Instruct/", save_dir);
 
     let message = r#"
     {
@@ -92,7 +96,7 @@ async fn qwen2_5vl_stream() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = Qwen2_5VLGenerateModel::init(model_path, None, None)?;
+    let mut model = Qwen2_5VLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
diff --git a/tests/test_qwen3vl.rs b/tests/test_qwen3vl.rs
index f7a5ce7..f52d662 100644
--- a/tests/test_qwen3vl.rs
+++ b/tests/test_qwen3vl.rs
@@ -9,7 +9,9 @@ use rocket::futures::StreamExt;
 fn qwen3vl_generate() -> Result<()> {
     // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda,ffmpeg qwen3vl_generate -r -- --nocapture
 
-    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen3-VL-2B-Instruct/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Qwen/Qwen3-VL-2B-Instruct/", save_dir);
 
     let message = r#"
     {
@@ -36,7 +38,7 @@ fn qwen3vl_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut qwen3vl = Qwen3VLGenerateModel::init(model_path, None, None)?;
+    let mut qwen3vl = Qwen3VLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
@@ -58,7 +60,9 @@ fn qwen3vl_generate() -> Result<()> {
 async fn qwen3vl_stream() -> Result<()> {
     // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda,ffmpeg qwen3vl_stream -r -- --nocapture
 
-    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen3-VL-2B-Instruct/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Qwen/Qwen3-VL-2B-Instruct/", save_dir);
 
     let message = r#"
     {
@@ -85,7 +89,7 @@ async fn qwen3vl_stream() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut qwen3vl = Qwen3VLGenerateModel::init(model_path, None, None)?;
+    let mut qwen3vl = Qwen3VLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
diff --git a/tests/test_rmbg2_0.rs b/tests/test_rmbg2_0.rs
index ef75833..346d8ab 100644
--- a/tests/test_rmbg2_0.rs
+++ b/tests/test_rmbg2_0.rs
@@ -8,7 +8,9 @@ use anyhow::Result;
 fn rmbg2_0_generate() -> Result<()> {
     // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda rmbg2_0_generate -r -- --nocapture
 
-    let model_path = "/home/jhq/huggingface_model/AI-ModelScope/RMBG-2.0/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/AI-ModelScope/RMBG-2.0/", save_dir);
 
     let message = r#"
     {
@@ -31,7 +33,7 @@ fn rmbg2_0_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let model = RMBG2_0Model::init(model_path, None, None)?;
+    let model = RMBG2_0Model::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
diff --git a/tests/test_robo_brain.rs b/tests/test_robo_brain.rs
index 2352c14..ecf1cde 100644
--- a/tests/test_robo_brain.rs
+++ b/tests/test_robo_brain.rs
@@ -8,7 +8,9 @@ use anyhow::Result;
 fn robo_brain_generate() -> Result<()> {
     // test with cuda: RUST_BACKTRACE=1 cargo test -F cuda robo_brain_generate -r -- --nocapture
 
-    let model_path = "/home/jhq/huggingface_model/BAAI/RoboBrain2.0-3B/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/BAAI/RoboBrain2.0-3B/", save_dir);
 
     let message = r#"
     {
@@ -28,7 +30,7 @@ fn robo_brain_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut model = Qwen2_5VLGenerateModel::init(model_path, None, None)?;
+    let mut model = Qwen2_5VLGenerateModel::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
diff --git a/tests/test_voxcpm.rs b/tests/test_voxcpm.rs
index 783fb7c..c45464b 100644
--- a/tests/test_voxcpm.rs
+++ b/tests/test_voxcpm.rs
@@ -13,7 +13,9 @@ use anyhow::{Ok, Result};
 #[test]
 fn voxcpm_use_message_generate() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda voxcpm_use_message_generate -r -- --nocapture
-    let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM-0.5B/", save_dir);
     let message = r#"
     {
         "model": "voxcpm",
@@ -40,7 +42,7 @@ fn voxcpm_use_message_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut voxcpm_generate = VoxCPMGenerate::init(model_path, None, None)?;
+    let mut voxcpm_generate = VoxCPMGenerate::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
@@ -59,10 +61,12 @@ fn voxcpm_use_message_generate() -> Result<()> {
 #[test]
 fn voxcpm_generate() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda voxcpm_generate -r -- --nocapture
-    let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM-0.5B/", save_dir);
 
     let i_start = Instant::now();
-    let mut voxcpm_generate = VoxCPMGenerate::init(model_path, None, None)?;
+    let mut voxcpm_generate = VoxCPMGenerate::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
@@ -106,8 +110,10 @@ fn voxcpm_generate() -> Result<()> {
 
 #[test]
 fn voxcpm_tokenizer() -> Result<()> {
-    let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
-    let tokenizer = SingleChineseTokenizer::new(model_path)?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM-0.5B/", save_dir);
+    let tokenizer = SingleChineseTokenizer::new(&model_path)?;
     let ids = tokenizer.encode("你好啊，你吃饭了吗".to_string())?;
     println!("ids: {:?}", ids);
     Ok(())
diff --git a/tests/test_voxcpm1_5.rs b/tests/test_voxcpm1_5.rs
index da1111a..4c7f452 100644
--- a/tests/test_voxcpm1_5.rs
+++ b/tests/test_voxcpm1_5.rs
@@ -13,7 +13,9 @@ use anyhow::{Ok, Result};
 #[test]
 fn voxcpm1_5_use_message_generate() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda voxcpm1_5_use_message_generate -r -- --nocapture
-    let model_path = "/home/jhq/huggingface_model/OpenBMB/VoxCPM1.5/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM1.5/", save_dir);
     let message = r#"
     {
         "model": "voxcpm1.5",
@@ -40,7 +42,7 @@ fn voxcpm1_5_use_message_generate() -> Result<()> {
     "#;
     let mes: ChatCompletionParameters = serde_json::from_str(message)?;
     let i_start = Instant::now();
-    let mut voxcpm_generate = VoxCPMGenerate::init(model_path, None, None)?;
+    let mut voxcpm_generate = VoxCPMGenerate::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
@@ -58,10 +60,12 @@ fn voxcpm1_5_use_message_generate() -> Result<()> {
 #[test]
 fn voxcpm1_5_generate() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda voxcpm1_5_generate -r -- --nocapture
-    let model_path = "/home/jhq/huggingface_model/OpenBMB/VoxCPM1.5/";
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM1.5/", save_dir);
 
     let i_start = Instant::now();
-    let mut voxcpm_generate = VoxCPMGenerate::init(model_path, None, None)?;
+    let mut voxcpm_generate = VoxCPMGenerate::init(&model_path, None, None)?;
     let i_duration = i_start.elapsed();
     println!("Time elapsed in load model is: {:?}", i_duration);
 
@@ -106,8 +110,10 @@ fn voxcpm1_5_generate() -> Result<()> {
 #[test]
 fn voxcpm1_5_tokenizer() -> Result<()> {
     // RUST_BACKTRACE=1 cargo test -F cuda voxcpm1_5_tokenizer -r -- --nocapture
-    let model_path = "/home/jhq/huggingface_model/OpenBMB/VoxCPM1.5/";
-    let tokenizer = SingleChineseTokenizer::new(model_path)?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM1.5/", save_dir);
+    let tokenizer = SingleChineseTokenizer::new(&model_path)?;
     let ids = tokenizer.encode("你好啊，你吃饭了吗".to_string())?;
     println!("ids: {:?}", ids);
     Ok(())
diff --git a/tests/weight_test.rs b/tests/weight_test.rs
index bbfa740..730d1f3 100644
--- a/tests/weight_test.rs
+++ b/tests/weight_test.rs
@@ -7,8 +7,10 @@ use candle_nn::VarBuilder;
 
 #[test]
 fn minicpm4_weight() -> Result<()> {
-    let model_path = "/home/jhq/huggingface_model/OpenBMB/MiniCPM4-0.5B/";
-    let model_list = find_type_files(model_path, "safetensors")?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/MiniCPM4-0.5B/", save_dir);
+    let model_list = find_type_files(&model_path, "safetensors")?;
     let device = Device::Cpu;
     for m in model_list {
         let weights = safetensors::load(m, &device)?;
@@ -23,8 +25,10 @@ fn minicpm4_weight() -> Result<()> {
 
 #[test]
 fn voxcpm_weight() -> Result<()> {
-    let model_path = "/home/jhq/huggingface_model/openbmb/VoxCPM-0.5B/";
-    let model_list = find_type_files(model_path, "pth")?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM-0.5B/", save_dir);
+    let model_list = find_type_files(&model_path, "pth")?;
     println!("model_list: {:?}", model_list);
     let dev = get_device(None);
     let mut dict_to_hashmap = HashMap::new();
@@ -48,8 +52,10 @@ fn voxcpm_weight() -> Result<()> {
 
 #[test]
 fn voxcpm1_5_weight() -> Result<()> {
-    let model_path = "/home/jhq/huggingface_model/OpenBMB/VoxCPM1.5/";
-    let model_list = find_type_files(model_path, "pth")?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/OpenBMB/VoxCPM1.5/", save_dir);
+    let model_list = find_type_files(&model_path, "pth")?;
     println!("model_list: {:?}", model_list);
     // let dev = get_device(None);
     let mut dict_to_hashmap = HashMap::new();
@@ -68,8 +74,10 @@ fn voxcpm1_5_weight() -> Result<()> {
 
 #[test]
 fn qwen3vl_weight() -> Result<()> {
-    let model_path = "/home/jhq/huggingface_model/Qwen/Qwen3-VL-4B-Instruct/";
-    let model_list = find_type_files(model_path, "safetensors")?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Qwen/Qwen3-VL-4B-Instruct/", save_dir);
+    let model_list = find_type_files(&model_path, "safetensors")?;
 
     let device = Device::Cpu;
     for m in &model_list {
@@ -84,8 +92,10 @@ fn qwen3vl_weight() -> Result<()> {
 
 #[test]
 fn deepseekocr_weight() -> Result<()> {
-    let model_path = "/home/jhq/huggingface_model/deepseek-ai/DeepSeek-OCR/";
-    let model_list = find_type_files(model_path, "safetensors")?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/deepseek-ai/DeepSeek-OCR/", save_dir);
+    let model_list = find_type_files(&model_path, "safetensors")?;
 
     let device = Device::Cpu;
     for m in &model_list {
@@ -103,8 +113,10 @@ fn deepseekocr_weight() -> Result<()> {
 
 #[test]
 fn hunyuanocr_weight() -> Result<()> {
-    let model_path = "/home/jhq/huggingface_model/Tencent-Hunyuan/HunyuanOCR/";
-    let model_list = find_type_files(model_path, "safetensors")?;
+    let save_dir =
+        aha::utils::get_default_save_dir().ok_or(anyhow::anyhow!("Failed to get save dir"))?;
+    let model_path = format!("{}/Tencent-Hunyuan/HunyuanOCR/", save_dir);
+    let model_list = find_type_files(&model_path, "safetensors")?;
 
     let device = Device::Cpu;
     for m in &model_list {