diff --git a/Dockerfile b/Dockerfile
index 231a3be..9b37f14 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -35,7 +35,8 @@ ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 RUN pip install --upgrade pip
 
 # Install ARM64 CUDA 13 wheels for PyTorch stack
-RUN pip install --no-cache-dir torch torchvision torchaudio \
+# Pinned to the versions validated on GB10 (CUDA graphs on SM121).
+RUN pip install --no-cache-dir torch==2.12.1 torchvision==0.27.1 torchaudio==2.11.0 \
     --index-url https://download.pytorch.org/whl/cu130
 
 # Install faster-qwen3-tts and server dependencies
diff --git a/README.md b/README.md
index 8cb994e..69f27fb 100644
--- a/README.md
+++ b/README.md
@@ -19,8 +19,8 @@ One Docker image covers all four backends, with two semantic tag aliases:
 
 | Tag | Use for |
 |---|---|
-| `:latest` / `:v6` | VoiceClone, VoiceDesign, CustomVoice |
-| `:latest-streaming` / `:v6-streaming` | Streaming VoiceClone |
+| `:latest` / `:v6` / `:v6.11` | VoiceClone, VoiceDesign, CustomVoice |
+| `:latest-streaming` / `:v6-streaming` / `:v6.11-streaming` | Streaming VoiceClone |
 
 Both tags point to the same image — the `-streaming` suffix is a semantic convention so compose files and version pins are unambiguous.
 
@@ -381,6 +381,7 @@ The first request after container startup can be slower because CUDA graph captu
 ### v6.11 — 2026-09-21
 **Feature: Language Aliases for VoiceClone Voices**
 - OpenAI-compatible clients that send a route alias such as `FR`, `FR_F` or `FR_M` in the `voice` field now get a configured speaker in that language (case-insensitive) instead of silently falling back to the default English voice. Unknown aliases still return the default voice or a 400 as before.
+- **Reproducible builds:** the PyTorch stack is pinned to the versions validated on GB10 (`torch==2.12.1`, `torchvision==0.27.1`, `torchaudio==2.11.0`) so a rebuild no longer picks up an untested PyTorch release.
 
 ### v6.10.1 — 2026-09-21
 **Fix: Fresh `docker build` Crashes on Model Load**