diff --git a/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml b/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml
index 02b52bc6c4..da7800ba66 100644
--- a/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml
+++ b/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml
@@ -36,7 +36,7 @@ spec:
   # Native upstream GTT-leak mitigation in LLMKube >=0.9.10; lifetime includes
   # startup and model load before this Vulkan embedder is recycled.
   maxPodLifetimeSeconds: 86400
-  image: ghcr.io/ggml-org/llama.cpp:server-vulkan@sha256:f8ec293aa54a6d42fa60bbfe3f37f6d4878d9df0d2e96e80f5d2dc514bcea182
+  image: ghcr.io/ggml-org/llama.cpp:server-vulkan@sha256:938a0fed6f4f83d00e0c2bac0c9ddfef343b2e8b2fd493a0abc442c4edb346ef
   # Two slots are the Vulkan-safe floor; more interleaving did not improve
   # throughput on this compute-bound iGPU. VMCP has its own embedder now.
   parallelSlots: 2 # Vulkan warmup hangs at parallel=1 (llama.cpp #24307)
diff --git a/kubernetes/apps/ai/llmkube/models/qwen35-2b.yaml b/kubernetes/apps/ai/llmkube/models/qwen35-2b.yaml
index e3a995c462..cb3ed2fe3f 100644
--- a/kubernetes/apps/ai/llmkube/models/qwen35-2b.yaml
+++ b/kubernetes/apps/ai/llmkube/models/qwen35-2b.yaml
@@ -33,7 +33,7 @@ spec:
   modelRef: qwen35-2b
   # Must be recent enough for Qwen3.5's GDN (Gated Delta Networks) architecture;
   # verified to load `qwen35` and run GDN fully on Vulkan (no CPU fallback).
-  image: ghcr.io/ggml-org/llama.cpp:server-vulkan@sha256:f8ec293aa54a6d42fa60bbfe3f37f6d4878d9df0d2e96e80f5d2dc514bcea182
+  image: ghcr.io/ggml-org/llama.cpp:server-vulkan@sha256:938a0fed6f4f83d00e0c2bac0c9ddfef343b2e8b2fd493a0abc442c4edb346ef
   parallelSlots: 2 # Vulkan warmup hangs at parallel=1 (llama.cpp #24307)
   # llama.cpp splits contextSize across parallelSlots, so 32768/2 = 16384 per slot.
   # GDN keeps KV tiny, so long ctx is cheap; 16K per slot is plenty for offload.
diff --git a/kubernetes/apps/ai/llmkube/models/vmcp-embedding.yaml b/kubernetes/apps/ai/llmkube/models/vmcp-embedding.yaml
index 547095f0fa..b951f767f0 100644
--- a/kubernetes/apps/ai/llmkube/models/vmcp-embedding.yaml
+++ b/kubernetes/apps/ai/llmkube/models/vmcp-embedding.yaml
@@ -42,7 +42,7 @@ spec:
   # Native upstream GTT-leak mitigation in LLMKube >=0.9.10; lifetime includes
   # startup and model load before this Vulkan embedder is recycled.
   maxPodLifetimeSeconds: 86400
-  image: ghcr.io/ggml-org/llama.cpp:server-vulkan@sha256:f8ec293aa54a6d42fa60bbfe3f37f6d4878d9df0d2e96e80f5d2dc514bcea182
+  image: ghcr.io/ggml-org/llama.cpp:server-vulkan@sha256:938a0fed6f4f83d00e0c2bac0c9ddfef343b2e8b2fd493a0abc442c4edb346ef
   # Vulkan floor is 2 — warmup hangs at parallel=1 (llama.cpp #24307). The iGPU
   # is compute-bound, so more slots add interleave, not throughput.
   parallelSlots: 2