diff --git a/sparks/inference/compose.yaml b/sparks/inference/compose.yaml
index 1472e3ed..cf9d59a8 100644
--- a/sparks/inference/compose.yaml
+++ b/sparks/inference/compose.yaml
@@ -7,7 +7,7 @@
 # model that fits in one node's budget.
 services:
   vllm:
-    image: vllm/vllm-openai:latest@sha256:61fc8a896b0a4fbbbdc063bc4b0dbc25ce98e02b5050c24aeb7830ac02039b14
+    image: vllm/vllm-openai:latest@sha256:c2914767605584b6d8f45686b82de173ecc99e781897aa3d0a66dacd72c51ae1
     container_name: vllm-qwen
     restart: unless-stopped
     # Host networking keeps the endpoint off the docker bridge, which the