diff --git a/.buildkite/hooks/pre-command b/.buildkite/hooks/pre-command
index 95483e7ae..681ac0db7 100644
--- a/.buildkite/hooks/pre-command
+++ b/.buildkite/hooks/pre-command
@@ -143,9 +143,9 @@ pipeline_add_env() {
 # Try to skip unaffected tests on pre-submits.
 # BUILDKITE_PIPELINE_ID is set only on the pipeline upload step.
 if [[ -n "${BUILDKITE_PIPELINE_ID}" && "${BUILDKITE_BRANCH}" =~ ^test/ ]]; then
-  # Skip load-all-images tests if the change doesn't affect images.
+  # Skip load-all-test-images tests if the change doesn't affect images.
   if ./tools/builkite-check-paths.sh images/ tools/images.mk Makefile; then
-    echo "Skip load-all-images tests"
+    echo "Skip load-all-test-images tests"
     pipeline_add_env SKIP_LOADING_IMAGES 1
   fi
 fi
diff --git a/.buildkite/pipeline.yaml b/.buildkite/pipeline.yaml
index 7fc2ab296..8af8953df 100644
--- a/.buildkite/pipeline.yaml
+++ b/.buildkite/pipeline.yaml
@@ -215,13 +215,13 @@ steps:
   - <<: *common
     label: ":docker: Images (x86_64)"
     if: build.env("SKIP_LOADING_IMAGES") == null && build.env("STAGED_BINARIES") == null
-    command: make ARCH=x86_64 load-all-images
+    command: make ARCH=x86_64 load-all-test-images
     agents:
       arch: "amd64"
   - <<: *common
     label: ":docker: Images (aarch64)"
     if: build.env("SKIP_LOADING_IMAGES") == null && build.env("STAGED_BINARIES") == null
-    command: make ARCH=aarch64 load-all-images
+    command: make ARCH=aarch64 load-all-test-images
     agents:
       arch: "arm64"
 
diff --git a/.buildkite/release.yaml b/.buildkite/release.yaml
index fe674873d..f8c0217ca 100644
--- a/.buildkite/release.yaml
+++ b/.buildkite/release.yaml
@@ -2,7 +2,7 @@ agents:
   queue: release
 _templates:
   common: &common
-    timeout_in_minutes: 120
+    timeout_in_minutes: 180
     retry:
       automatic:
         - exit_status: -1
diff --git a/images/README.md b/images/README.md
index 0b064446e..5784341e9 100644
--- a/images/README.md
+++ b/images/README.md
@@ -29,7 +29,7 @@ image infrastructure.
 
 The continuous integration system can either take fine-grained dependencies on
 single images via individual `load` targets, or pull all images via a single
-`load-all-images` invocation.
+`load-all-test-images` invocation.
 
 ## Adding new images
 
diff --git a/tools/images.mk b/tools/images.mk
index f15e31680..bcd4474b9 100644
--- a/tools/images.mk
+++ b/tools/images.mk
@@ -41,10 +41,12 @@ endif
 # runsc/testutil.MangleImage. Names are mangled in this way to ensure that all
 # tests are using locally-defined images (that are consistent and idempotent).
 REMOTE_IMAGE_PREFIX ?= gcr.io/gvisor-presubmit
-LOCAL_IMAGE_PREFIX ?= gvisor.dev/images
-ALL_IMAGES   := $(subst /,_,$(subst images/,,$(shell find images/ -name Dockerfile -o -name Dockerfile.$(ARCH) | xargs -n 1 dirname | uniq)))
-SUB_IMAGES   := $(foreach image,$(ALL_IMAGES),$(if $(findstring _,$(image)),$(image),))
-IMAGE_GROUPS := $(sort $(foreach image,$(SUB_IMAGES),$(firstword $(subst _, ,$(image)))))
+LOCAL_IMAGE_PREFIX  ?= gvisor.dev/images
+ALL_IMAGES          := $(subst /,_,$(subst images/,,$(shell find images/ -name Dockerfile -o -name Dockerfile.$(ARCH) | xargs -n 1 dirname | uniq)))
+NON_TEST_IMAGES     := gpu/ollama/bench
+TEST_IMAGES         := $(subst /,_,$(subst images/,,$(shell find images/ -name Dockerfile -o -name Dockerfile.$(ARCH) | xargs -n 1 dirname | uniq | grep -v $(NON_TEST_IMAGES))))
+SUB_IMAGES          := $(foreach image,$(ALL_IMAGES),$(if $(findstring _,$(image)),$(image),))
+IMAGE_GROUPS        := $(sort $(foreach image,$(SUB_IMAGES),$(firstword $(subst _, ,$(image)))))
 
 define expand_group =
 load-$(1): $$(patsubst $(1)_%, load-$(1)_%, $$(filter $(1)_%,$$(ALL_IMAGES)))
@@ -60,14 +62,26 @@ list-all-images: ## List all images.
 	@for image in $(ALL_IMAGES); do echo $${image}; done
 .PHONY: list-all-images
 
+list-all-test-images: ## List all test images.
+	@for image in $(TEST_IMAGES); do echo $${image}; done
+.PHONY: list-all-test-images
+
 load-all-images: ## Load all images.
 load-all-images: $(patsubst %,load-%,$(ALL_IMAGES))
 .PHONY: load-all-images
 
+load-all-test-images: ## Load all test images.
+load-all-test-images: $(patsubst %,load-%,$(TEST_IMAGES))
+.PHONY: load-all-test-images
+
 push-all-images: ## Push all images.
 push-all-images: $(patsubst %,push-%,$(ALL_IMAGES))
 .PHONY: push-all-images
 
+push-all-test-images: ## Push all images.
+push-all-test-images: $(patsubst %,push-%,$(TEST_IMAGES))
+.PHONY: push-all-test-images
+
 # path and dockerfile are used to extract the relevant path and dockerfile
 # (depending on what's available for the given architecture).
 path = images/$(subst _,/,$(1))