diff --git a/kischdle/llmux/scripts/download_models.sh b/kischdle/llmux/scripts/download_models.sh
index facefad..e6da9dc 100755
--- a/kischdle/llmux/scripts/download_models.sh
+++ b/kischdle/llmux/scripts/download_models.sh
@@ -1,15 +1,26 @@
 #!/bin/bash
 # Download all model weights for llmux.
 # Run as user llm: bash scripts/download_models.sh
-# Requires: pip install huggingface_hub
 # Requires: HuggingFace token at ~/.cache/huggingface/token for gated models
 
 set -euo pipefail
 
+# Use llm user's venv for huggingface-cli and python
+VENV="${LLMUX_VENV:-$HOME/.venv-pytorch}"
+HF_CLI="${VENV}/bin/huggingface-cli"
+PYTHON="${VENV}/bin/python"
+
+if [ ! -x "$HF_CLI" ]; then
+    echo "ERROR: huggingface-cli not found at $HF_CLI"
+    echo "Install with: ${VENV}/bin/pip install huggingface_hub"
+    exit 1
+fi
+
 MODELS_DIR="${LLMUX_MODELS_DIR:-$HOME/.local/share/llmux_pod/models}"
 mkdir -p "$MODELS_DIR"
 
 echo "=== Downloading models to $MODELS_DIR ==="
+echo "Using: $HF_CLI"
 
 download_hf() {
     local repo="$1"
@@ -19,14 +30,14 @@ download_hf() {
         return
     fi
     echo "Downloading: $repo"
-    huggingface-cli download "$repo" --cache-dir "$MODELS_DIR"
+    "$HF_CLI" download "$repo" --cache-dir "$MODELS_DIR"
 }
 
 download_hf_files() {
     local repo="$1"
     shift
     echo "Downloading specific files from: $repo"
-    huggingface-cli download "$repo" "$@" --cache-dir "$MODELS_DIR"
+    "$HF_CLI" download "$repo" "$@" --cache-dir "$MODELS_DIR"
 }
 
 # 1. Qwen3.5-9B-FP8
@@ -53,7 +64,7 @@ download_hf "CohereLabs/cohere-transcribe-03-2026" || \
 
 # 7. Chatterbox TTS
 echo "Downloading: Chatterbox TTS weights (auto-downloaded by library)"
-python3 -c "
+"$PYTHON" -c "
 from chatterbox.tts import ChatterboxTTS
 import os
 os.environ['CUDA_VISIBLE_DEVICES'] = ''