diff --git a/backends/webgpu/test/ops/test_quantized_linear.py b/backends/webgpu/test/ops/test_quantized_linear.py
index b1a033502b3..14fef6be33c 100644
--- a/backends/webgpu/test/ops/test_quantized_linear.py
+++ b/backends/webgpu/test/ops/test_quantized_linear.py
@@ -63,6 +63,13 @@ class Q4gswConfig:
     # (e.g. lvp); else shmem (K>=4096 or N>=2048) or register-tiled (SwiftShader
     # caps at 128). Same fp64 golden regardless of which kernel runs.
     Q4gswConfig("steel", 96, 2048, 256),  # steel-isolating (K<4096, N<2048)
+    # Same shape as "steel"; the .pte is dtype-independent, so this fixture feeds
+    # the f16-multiply steel kernel in a -DWGPU_BACKEND_STEEL_F16 build (goldened
+    # at a looser f16 tol in the native test). The default f32 build ignores it.
+    Q4gswConfig("steel_f16", 96, 2048, 256),  # f16-multiply steel (opt-in build)
+    # Partial M and N steel tiles under the f16 kernel; exercises f16 boundary
+    # masking (the exact-N "steel_f16" shape does not). N%8==0, steel-isolating.
+    Q4gswConfig("steel_f16_edge", 70, 1024, 136),  # f16 partial-tile (opt-in)
     Q4gswConfig("gate_proj_pf", 128, 2048, 8192),  # gate/up prefill (shmem via N)
     Q4gswConfig("down_proj_pf", 128, 8192, 2048),  # down prefill (shmem via K)
     Q4gswConfig("shmem_edge", 130, 4096, 2056),  # partial 32-tile bounds
diff --git a/backends/webgpu/test/test_webgpu_native.cpp b/backends/webgpu/test/test_webgpu_native.cpp
index dea4f26ff20..1e8034d0e1e 100644
--- a/backends/webgpu/test/test_webgpu_native.cpp
+++ b/backends/webgpu/test/test_webgpu_native.cpp
@@ -277,6 +277,14 @@ const Q4gswConfig kQ4gswConfigs[] = {
     // general-N robustness).
     // M>1: steel GEMM on a >=256-invocation device (K%16==0), else shmem/tiled.
     {"steel", 96, 2048, 256, 1e-4f, 1e-3f, true, false}, // steel-isolating
+#ifdef WGPU_BACKEND_STEEL_F16
+    // Same shape as "steel" run under the f16-multiply steel kernel; the f16
+    // rounding floor (~2.3e-4, uniform in K -- not an accumulate bug) needs a
+    // looser abs gate than the strict f32 1e-4. Opt-in build only.
+    {"steel_f16", 96, 2048, 256, 2.3e-4f, 1e-3f, true, false},
+    // Partial M and N steel tiles under the f16 kernel (f16 boundary masking).
+    {"steel_f16_edge", 70, 1024, 136, 2.3e-4f, 1e-3f, true, false},
+#endif
     {"gate_proj_pf", 128, 2048, 8192, 1e-4f, 1e-3f, true, false}, // shmem via N
     {"down_proj_pf", 128, 8192, 2048, 1e-3f, 1e-2f, true, false}, // shmem via K
     {"shmem_edge", 130, 4096, 2056, 1e-4f, 1e-3f, true, false}, // partial tiles