diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml
index f75d7812859e..3ba1b221ea80 100644
--- a/.github/workflows/main.yml
+++ b/.github/workflows/main.yml
@@ -678,6 +678,9 @@ jobs:
       matrix:
         feature: ["openvino", "onnx"]
         os: ["ubuntu-latest", "windows-latest"]
+        include:
+          - os: windows-latest
+            feature: winml
     name: Test wasi-nn (${{ matrix.feature }}, ${{ matrix.os }})
     runs-on: ${{ matrix.os }}
     needs: determine
@@ -692,6 +695,15 @@ jobs:
     - uses: abrown/install-openvino-action@v8
       if: runner.arch == 'X64'
 
+    # Install WinML for testing wasi-nn WinML backend. WinML is only available
+    # on Windows clients and Windows Server with desktop experience enabled.
+    # GitHub Actions Window Server image doesn't have desktop experience
+    # enabled, so we download the standalone library from ONNX Runtime project.
+    - uses: nuget/setup-nuget@v2
+      if: (matrix.os == 'windows-latest') && (matrix.feature == 'winml')
+    - run: nuget install Microsoft.AI.MachineLearning
+      if: (matrix.os == 'windows-latest') && (matrix.feature == 'winml')
+
     # Install Rust targets.
     - run: rustup target add wasm32-wasi
 
diff --git a/crates/test-programs/src/bin/nn_image_classification_onnx.rs b/crates/test-programs/src/bin/nn_image_classification_onnx.rs
index 30d20330f39f..df8f961c7516 100644
--- a/crates/test-programs/src/bin/nn_image_classification_onnx.rs
+++ b/crates/test-programs/src/bin/nn_image_classification_onnx.rs
@@ -8,10 +8,13 @@ pub fn main() -> Result<()> {
         .expect("the model file to be mapped to the fixture directory");
     let graph =
         GraphBuilder::new(GraphEncoding::Onnx, ExecutionTarget::CPU).build_from_bytes([&model])?;
-    let tensor = fs::read("fixture/tensor.bgr")
+    let tensor = fs::read("fixture/000000062808.rgb")
         .expect("the tensor file to be mapped to the fixture directory");
     let results = classify(graph, tensor)?;
     let top_five = &sort_results(&results)[..5];
+    // 963 is meat loaf, meatloaf.
+    // https://github.com/onnx/models/blob/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/synset.txt#L963
+    assert_eq!(top_five[0].class_id(), 963);
     println!("found results, sorted top 5: {:?}", top_five);
     Ok(())
 }
diff --git a/crates/test-programs/src/nn.rs b/crates/test-programs/src/nn.rs
index cf25e4b8979e..b7cc0e562230 100644
--- a/crates/test-programs/src/nn.rs
+++ b/crates/test-programs/src/nn.rs
@@ -42,12 +42,15 @@ pub fn classify(graph: Graph, tensor: Vec<u8>) -> Result<Vec<f32>> {
 /// placing the match probability for each class at the index for that class
 /// (the probability of class `N` is stored at `probabilities[N]`).
 pub fn sort_results(probabilities: &[f32]) -> Vec<InferenceResult> {
+    let probabilities_iter = probabilities.iter();
+
     // It is unclear why the MobileNet output indices are "off by one" but the
     // `.skip(1)` below seems necessary to get results that make sense (e.g. 763
     // = "revolver" vs 762 = "restaurant").
-    let mut results: Vec<InferenceResult> = probabilities
-        .iter()
-        .skip(1)
+    #[cfg(feature = "openvino")]
+    probabilities_iter.skip(1);
+
+    let mut results: Vec<InferenceResult> = probabilities_iter
         .enumerate()
         .map(|(c, p)| InferenceResult(c, *p))
         .collect();
diff --git a/crates/wasi-nn/Cargo.toml b/crates/wasi-nn/Cargo.toml
index 1b8dc95d99d6..43dba88519c8 100644
--- a/crates/wasi-nn/Cargo.toml
+++ b/crates/wasi-nn/Cargo.toml
@@ -50,7 +50,7 @@ wasi-common = { workspace = true, features = ["sync"] }
 wasmtime = { workspace = true, features = ["cranelift"] }
 
 [features]
-default = ["openvino"]
+default = ["openvino", "winml"]
 # openvino is available on all platforms, it requires openvino installed.
 openvino = ["dep:openvino"]
 # onnx is available on all platforms.
diff --git a/crates/wasi-nn/src/testing.rs b/crates/wasi-nn/src/testing.rs
index 51b543e15a95..da0360898da5 100644
--- a/crates/wasi-nn/src/testing.rs
+++ b/crates/wasi-nn/src/testing.rs
@@ -8,7 +8,12 @@
 
 #[allow(unused_imports)]
 use anyhow::{anyhow, Context, Result};
-use std::{env, fs, path::Path, path::PathBuf, process::Command, sync::Mutex};
+use std::{
+    env, fs,
+    path::{Path, PathBuf},
+    process::Command,
+    sync::Mutex,
+};
 
 #[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
 use windows::AI::MachineLearning::{LearningModelDevice, LearningModelDeviceKind};
@@ -50,13 +55,12 @@ pub fn check() -> Result<()> {
     #[cfg(feature = "openvino")]
     check_openvino_artifacts_are_available()?;
 
-    #[cfg(feature = "onnx")]
+    #[cfg(any(feature = "onnx", all(feature = "winml", target_os = "windows")))]
     check_onnx_artifacts_are_available()?;
 
-    #[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
+    #[cfg(all(feature = "winml", target_os = "windows"))]
     {
         check_winml_is_available()?;
-        check_winml_artifacts_are_available()?;
     }
     Ok(())
 }
@@ -108,7 +112,7 @@ fn check_openvino_artifacts_are_available() -> Result<()> {
     Ok(())
 }
 
-#[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
+#[cfg(all(feature = "winml", target_os = "windows"))]
 fn check_winml_is_available() -> Result<()> {
     match std::panic::catch_unwind(|| {
         println!(
@@ -121,27 +125,19 @@ fn check_winml_is_available() -> Result<()> {
     }
 }
 
-#[cfg(feature = "onnx")]
+#[cfg(any(feature = "onnx", all(feature = "winml", target_os = "windows")))]
 fn check_onnx_artifacts_are_available() -> Result<()> {
     let _exclusively_retrieve_artifacts = ARTIFACTS.lock().unwrap();
 
-    const OPENVINO_BASE_URL: &str =
-        "https://github.com/intel/openvino-rs/raw/main/crates/openvino/tests/fixtures/mobilenet";
     const ONNX_BASE_URL: &str =
-        "https://github.com/onnx/models/raw/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/mobilenet/model/mobilenetv2-7.onnx?download=";
+        "https://github.com/onnx/models/raw/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/mobilenet/model/mobilenetv2-10.onnx?download=";
 
     let artifacts_dir = artifacts_dir();
     if !artifacts_dir.is_dir() {
         fs::create_dir(&artifacts_dir)?;
     }
 
-    for (from, to) in [
-        (
-            [OPENVINO_BASE_URL, "tensor-1x224x224x3-f32.bgr"].join("/"),
-            "tensor.bgr",
-        ),
-        (ONNX_BASE_URL.to_string(), "model.onnx"),
-    ] {
+    for (from, to) in [(ONNX_BASE_URL.to_string(), "model.onnx")] {
         let local_path = artifacts_dir.join(to);
         if !local_path.is_file() {
             download(&from, &local_path).with_context(|| "unable to retrieve test artifact")?;
@@ -149,31 +145,14 @@ fn check_onnx_artifacts_are_available() -> Result<()> {
             println!("> using cached artifact: {}", local_path.display())
         }
     }
-    Ok(())
-}
 
-#[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
-fn check_winml_artifacts_are_available() -> Result<()> {
-    let _exclusively_retrieve_artifacts = ARTIFACTS.lock().unwrap();
-    let artifacts_dir = artifacts_dir();
-    if !artifacts_dir.is_dir() {
-        fs::create_dir(&artifacts_dir)?;
-    }
-    const MODEL_URL: &str="https://github.com/onnx/models/raw/5faef4c33eba0395177850e1e31c4a6a9e634c82/vision/classification/mobilenet/model/mobilenetv2-12.onnx";
-    for (from, to) in [(MODEL_URL, "model.onnx")] {
-        let local_path = artifacts_dir.join(to);
-        if !local_path.is_file() {
-            download(&from, &local_path).with_context(|| "unable to retrieve test artifact")?;
-        } else {
-            println!("> using cached artifact: {}", local_path.display())
-        }
-    }
-    // kitten.rgb is converted from https://github.com/microsoft/Windows-Machine-Learning/blob/master/SharedContent/media/kitten_224.png?raw=true.
-    let tensor_path = env::current_dir()?
+    // Copy image from source tree to artifact directory.
+    let image_path = env::current_dir()?
         .join("tests")
         .join("fixtures")
-        .join("kitten.rgb");
-    fs::copy(tensor_path, artifacts_dir.join("kitten.rgb"))?;
+        .join("000000062808.rgb");
+    let dest_path = artifacts_dir.join("000000062808.rgb");
+    fs::copy(&image_path, &dest_path)?;
     Ok(())
 }
 
diff --git a/crates/wasi-nn/tests/all.rs b/crates/wasi-nn/tests/all.rs
index 06be156079d6..2c41f17db2e3 100644
--- a/crates/wasi-nn/tests/all.rs
+++ b/crates/wasi-nn/tests/all.rs
@@ -97,10 +97,10 @@ fn nn_image_classification_named() {
 #[cfg_attr(not(all(feature = "winml", target_os = "windows")), ignore)]
 #[test]
 fn nn_image_classification_winml() {
-    #[cfg(feature = "winml")]
+    #[cfg(all(feature = "winml", target_os = "windows"))]
     {
         let backend = Backend::from(backend::winml::WinMLBackend::default());
-        run(NN_IMAGE_CLASSIFICATION_WINML, backend, true).unwrap()
+        run(NN_IMAGE_CLASSIFICATION_ONNX, backend, true).unwrap()
     }
 }
 
diff --git a/crates/wasi-nn/tests/fixtures/000000062808.rgb b/crates/wasi-nn/tests/fixtures/000000062808.rgb
new file mode 100644
index 000000000000..ac79a904fb92
Binary files /dev/null and b/crates/wasi-nn/tests/fixtures/000000062808.rgb differ
diff --git a/crates/wasi-nn/tests/fixtures/kitten.rgb b/crates/wasi-nn/tests/fixtures/kitten.rgb
deleted file mode 100644
index d39e9f9e20cb..000000000000
Binary files a/crates/wasi-nn/tests/fixtures/kitten.rgb and /dev/null differ
diff --git a/crates/wasi-nn/tests/fixtures/readme.md b/crates/wasi-nn/tests/fixtures/readme.md
new file mode 100644
index 000000000000..7001167b666c
--- /dev/null
+++ b/crates/wasi-nn/tests/fixtures/readme.md
@@ -0,0 +1,9 @@
+The original image of 000000062808.rgb is 000000062808.jpg from ImageNet
+database. It processed by following Python code with
+https://github.com/onnx/models/blob/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/imagenet_preprocess.py
+
+```
+image = mxnet.image.imread('000000062808.jpg')
+image = preprocess_mxnet(image)
+image.asnumpy().tofile('000000062808.rgb')
+```
\ No newline at end of file