diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 36fd221..8610290 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -82,7 +82,9 @@ jobs:
         # Excluded (require external services):
         # ✗ S3, GCS, Azure, MongoDB, Mnesia
         #
-        # Exclusions are configured in test/test_helper.exs
+        # Exclusions are configured in test/test_helper.exs. Wall-clock and
+        # memory benchmarks (:performance, :memory) are excluded there too and
+        # run in the non-blocking "Benchmarks" job below.
         # To run all tests locally:
         # mix test --include s3 --include gcs --include azure --include mongo --include mnesia
         run: mix test --trace
@@ -182,6 +184,42 @@ jobs:
           echo "Checking for undocumented public functions..."
           mix docs --warnings-as-errors
 
+  # Wall-clock and memory benchmarks. Shared runners vary in speed and load, so
+  # these report regressions without blocking merges: a failure shows in this
+  # job's log and annotations while the workflow stays green.
+  benchmarks:
+    name: Benchmarks (non-blocking)
+    runs-on: ubuntu-latest
+    continue-on-error: true
+    env:
+      MIX_ENV: test
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Elixir
+        uses: erlef/setup-beam@v1
+        with:
+          elixir-version: "1.19.5"
+          otp-version: "28.1"
+
+      - name: Restore dependencies cache
+        uses: actions/cache@v4
+        with:
+          path: |
+            deps
+            _build
+          key: ${{ runner.os }}-mix-benchmarks-${{ hashFiles('**/mix.lock') }}
+          restore-keys: |
+            ${{ runner.os }}-mix-benchmarks-
+
+      - name: Install dependencies
+        run: mix deps.get
+
+      - name: Run benchmark tests
+        run: mix test --only performance --only memory
+
   livebooks:
     name: Livebooks
     runs-on: ubuntu-latest
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 9e85b12..c11438f 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -85,8 +85,9 @@ jobs:
         run: mix compile --warnings-as-errors
 
       # Functional suite. Wall-clock and memory benchmarks (:performance,
-      # :memory) are noisy on shared runners when tests run in parallel; CI
-      # runs them serially (mix test --trace) on every push to main.
+      # :memory) are excluded by test/test_helper.exs; the flags below repeat
+      # that so the release gate does not depend on it. CI runs them in its
+      # non-blocking "Benchmarks" job.
       - name: Run tests
         run: mix test --exclude performance --exclude memory
 
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1e4ce90..2a3cace 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -47,6 +47,12 @@ docs that match measured behaviour. No public API was removed or renamed.
   say that Explorer streaming was planned for 1.3 and dropped.
 
 ### Tests
+- Wall-clock and memory benchmark tests (`:performance`, `:memory`) are
+  excluded from the default `mix test` run and from blocking CI. They failed
+  CI three times on shared runners with no code change (for example, an 8 MB
+  `from_tensor/2` took 6.3 s on one runner and 0.64 s locally). CI runs them
+  in a separate non-blocking "Benchmarks" job; run them locally with
+  `mix test --only performance --only memory`.
 - The streaming memory test measures the test process instead of the whole VM,
   so other async tests no longer move the result; the limit is now 2 MB.
 - New tests cover `stream_chunk_tensors/2` options (`:include_missing`,
diff --git a/test/test_helper.exs b/test/test_helper.exs
index c128f42..3ed2c21 100644
--- a/test/test_helper.exs
+++ b/test/test_helper.exs
@@ -13,6 +13,16 @@
 #
 # To run live cloud/Python tests locally:
 #   mix test --include s3 --include gcs --include azure --include mongo --include mnesia --include python
-ExUnit.configure(exclude: [:s3, :gcs, :azure, :mongo, :mnesia, :python, :python_fixtures])
+#
+# Wall-clock and memory benchmarks are also excluded by default:
+# - :performance - asserts elapsed time (e.g. an 8 MB conversion under 5 s)
+# - :memory      - asserts memory growth while streaming
+# Their results depend on how loaded the machine is, so on shared CI runners
+# they fail without any code change. CI runs them in the non-blocking
+# "Benchmarks" job. Run them locally with:
+#   mix test --only performance --only memory
+ExUnit.configure(
+  exclude: [:s3, :gcs, :azure, :mongo, :mnesia, :python, :python_fixtures, :performance, :memory]
+)
 
 ExUnit.start()