diff --git a/src/nn/linear.rs b/src/nn/linear.rs
index 7bc5c92..8cf8369 100644
--- a/src/nn/linear.rs
+++ b/src/nn/linear.rs
@@ -35,8 +35,13 @@ impl Linear {
     /// its own shift follows immediately.
     pub fn no_bias(in_features: usize, out_features: usize) -> Linear {
         Linear {
+            weight: Param::new(Tensor::kaiming_uniform(
+                &[out_features, in_features],
+                in_features,
+            )),
             bias: None,
-            ..Linear::new(in_features, out_features)
+            in_features,
+            out_features,
         }
     }
 
diff --git a/tests/linear.rs b/tests/linear.rs
new file mode 100644
index 0000000..93edbe6
--- /dev/null
+++ b/tests/linear.rs
@@ -0,0 +1,65 @@
+//! Linear initialization must consume only the random samples its parameters need.
+
+use fastnn::prelude::*;
+
+#[test]
+fn no_bias_uses_only_weight_initialization_draws() {
+    for seed in [0, 7, 42] {
+        manual_seed(seed);
+        let layer = Linear::no_bias(2, 3);
+        let next = Tensor::rand(&[4]).to_vec();
+
+        manual_seed(seed);
+        let weight = Tensor::kaiming_uniform(&[3, 2], 2);
+        let expected_next = Tensor::rand(&[4]).to_vec();
+
+        assert!(layer.bias.is_none());
+        assert_eq!(layer.weight.value().to_vec(), weight.to_vec());
+        assert_eq!(
+            next, expected_next,
+            "no_bias consumed samples beyond its weight for seed {seed}"
+        );
+    }
+}
+
+#[test]
+fn new_preserves_weight_then_bias_initialization() {
+    for seed in [0, 7, 42] {
+        manual_seed(seed);
+        let layer = Linear::new(2, 3);
+        let next = Tensor::rand(&[4]).to_vec();
+
+        manual_seed(seed);
+        let weight = Tensor::kaiming_uniform(&[3, 2], 2);
+        let bound = 1.0 / 2.0_f32.sqrt();
+        let bias = Tensor::uniform(&[3], -bound, bound);
+        let expected_next = Tensor::rand(&[4]).to_vec();
+
+        assert_eq!(layer.weight.value().to_vec(), weight.to_vec());
+        assert_eq!(layer.bias.as_ref().unwrap().value().to_vec(), bias.to_vec());
+        assert_eq!(next, expected_next);
+    }
+}
+
+#[test]
+fn no_bias_preserves_forward_shape_and_trainable_weight() {
+    let layer = Linear::no_bias(2, 3);
+    let input = Tensor::from_vec(vec![1.0, 2.0, -3.0, 4.0], &[2, 1, 2]);
+
+    let output = layer.forward(&input);
+    let expected = input
+        .reshape(&[2, 2])
+        .matmul_nt(&layer.weight.value())
+        .reshape(&[2, 1, 3]);
+
+    assert_eq!(layer.in_features(), 2);
+    assert_eq!(layer.out_features(), 3);
+    assert_eq!(output.shape(), &[2, 1, 3]);
+    assert_eq!(output.to_vec(), expected.to_vec());
+    assert_eq!(layer.named_parameters()[0].0, "weight");
+    assert_eq!(layer.parameters().len(), 1);
+    assert_eq!(layer.trainable_parameters().len(), 1);
+
+    output.sum().backward();
+    assert_eq!(layer.weight.grad().unwrap().shape(), &[3, 2]);
+}