diff --git a/Dockerfile b/Dockerfile
index a2cd616..9892895 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -16,9 +16,10 @@ COPY src/ src/
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install .
 # Voice dependencies (faster-whisper, Kokoro TTS, soundfile) — activated at runtime via VOICE_ENABLED
-# Cache mount keeps wheels across builds so the heavy torch/CUDA deps aren't re-downloaded every time.
+# Install CPU-only torch first so pip doesn't pull full CUDA wheels (~2 GB) for kokoro/transformers.
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install faster-whisper kokoro soundfile
+    pip install torch --index-url https://download.pytorch.org/whl/cpu \
+    && pip install faster-whisper kokoro soundfile
 
 # Build the fable-mcp wheel so it can be served for download
 COPY fable-mcp/ fable-mcp/