diff --git a/.env.example b/.env.example
index 9481f6a..2f9cbc2 100644
--- a/.env.example
+++ b/.env.example
@@ -33,3 +33,6 @@ DEEPL_API_URL="https://api-free.deepl.com/v2/translate"  # Use https://api.deepl
 
 # AI Vision Service (Ollama for Alt Text Generation)
 OLLAMA_URL="http://localhost:11434"  # Local Ollama server URL
+
+# HuggingFace Transformers Model Cache (for nutrition embedding models)
+TRANSFORMERS_CACHE="/var/cache/transformers"  # Must be writable by build and runtime user
diff --git a/package.json b/package.json
index 89dc636..4af2ea8 100644
--- a/package.json
+++ b/package.json
@@ -5,7 +5,7 @@
 	"type": "module",
 	"scripts": {
 		"dev": "vite dev",
-		"prebuild": "bash scripts/subset-emoji-font.sh && pnpm exec vite-node scripts/generate-mystery-verses.ts",
+		"prebuild": "bash scripts/subset-emoji-font.sh && pnpm exec vite-node scripts/generate-mystery-verses.ts && pnpm exec vite-node scripts/download-models.ts",
 		"build": "vite build",
 		"preview": "vite preview",
 		"check": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json",
diff --git a/scripts/download-models.ts b/scripts/download-models.ts
new file mode 100644
index 0000000..cc1589c
--- /dev/null
+++ b/scripts/download-models.ts
@@ -0,0 +1,17 @@
+/**
+ * Pre-downloads HuggingFace transformer models so they're cached for runtime.
+ * Run with: pnpm exec vite-node scripts/download-models.ts
+ */
+import { pipeline } from '@huggingface/transformers';
+
+const MODELS = [
+	'Xenova/all-MiniLM-L6-v2',
+	'Xenova/multilingual-e5-small',
+];
+
+for (const name of MODELS) {
+	console.log(`Downloading ${name}...`);
+	const p = await pipeline('feature-extraction', name, { dtype: 'q8' });
+	await p.dispose();
+	console.log(`  done`);
+}