diff --git a/crates/burn/src/ops/matmul.rs b/crates/burn/src/ops/matmul.rs
index 4dabc3cc..b3ae7339 100644
--- a/crates/burn/src/ops/matmul.rs
+++ b/crates/burn/src/ops/matmul.rs
@@ -68,9 +68,176 @@ pub fn clear_attention_cache() {
 }
 
 // ============================================================================
-// VNNI u8 MatVec fast path — 64 MACs per instruction
+// Compiled Linear Cache — O(k) replacing O(n_rows) for any weight matrix
 // ============================================================================
 //
+// For any linear layer y = W @ x, where W is [n_rows, n_cols]:
+//   1. Each row of W is assigned to one of 256 palette centroids (u8 index)
+//   2. At inference: compute k=256 centroid dot products with input x
+//   3. For each output row i: y[i] = centroid_outputs[assignment[i]]
+//
+// Cost: 256 × n_cols MACs + n_rows lookups (vs n_rows × n_cols MACs)
+// For gate_proj [3072, 1024]: 256K MACs vs 3.1M MACs = 12× fewer.
+//
+// Keyed by (n_rows, n_cols) — the weight matrix shape.
+
+/// A compiled linear layer: 256 centroids replace the full weight matrix.
+#[cfg(feature = "std")]
+#[derive(Clone)]
+pub struct CompiledLinear {
+    /// Centroid weight vectors: [k × n_cols] f32, row-major.
+    /// k=256 centroids, each of dimension n_cols.
+    pub centroids: Vec<f32>,
+    /// Number of centroids (palette size, typically 256).
+    pub k: usize,
+    /// Input dimension (n_cols of the original weight matrix).
+    pub n_cols: usize,
+    /// Output dimension (n_rows of the original weight matrix).
+    pub n_rows: usize,
+    /// Row assignment: for each of the n_rows output rows, which centroid it maps to.
+    pub assignments: Vec<u8>,
+}
+
+/// Global cache of compiled linear layers.
+/// Keyed by (n_rows, n_cols) — the original weight matrix shape.
+/// Multiple layers can share the same shape, so we use a Vec and match by registration order.
+#[cfg(feature = "std")]
+static LINEAR_CACHE: LazyLock<RwLock<Vec<CompiledLinear>>> =
+    LazyLock::new(|| RwLock::new(Vec::new()));
+
+/// Register a compiled linear layer.
+#[cfg(feature = "std")]
+pub fn register_compiled_linear(compiled: CompiledLinear) {
+    let mut cache = LINEAR_CACHE.write().unwrap();
+    cache.push(compiled);
+}
+
+/// Pop the next compiled linear for the given shape.
+/// Returns None if no matching table exists.
+/// This is FIFO — layers are consumed in registration order.
+#[cfg(feature = "std")]
+fn pop_compiled_linear(n_rows: usize, n_cols: usize) -> Option<CompiledLinear> {
+    let cache = LINEAR_CACHE.read().unwrap();
+    // Find first matching entry (don't pop — layers may be reused across batches)
+    cache.iter().find(|c| c.n_rows == n_rows && c.n_cols == n_cols).cloned()
+}
+
+/// Try to compute y = W @ x using compiled centroid matmul with VNNI acceleration.
+///
+/// Instead of n_rows × n_cols MACs:
+///   1. Quantize centroids to u8, input column to i8
+///   2. VNNI dot: 256 centroid × input dots at 64 MACs/instruction
+///   3. Dequantize i32 results back to f32 via scale factors
+///   4. Broadcast via palette assignment: out[i] = centroid_out[assignment[i]]
+///
+/// Returns true if compiled path was used.
+#[cfg(feature = "std")]
+fn try_compiled_linear<E: NdArrayElement>(
+    _lhs: &ndarray::ArrayView2<'_, E>,
+    _rhs: &ndarray::ArrayView2<'_, E>,
+    out: &mut ndarray::ArrayViewMut2<'_, E>,
+    m: usize,
+    k_dim: usize,
+    n: usize,
+) -> bool {
+    let compiled = match pop_compiled_linear(m, k_dim) {
+        Some(c) => c,
+        None => return false,
+    };
+
+    if compiled.assignments.len() < m || compiled.k == 0 {
+        return false;
+    }
+
+    let k = compiled.k;
+    let dim = compiled.n_cols.min(k_dim);
+
+    // Pre-quantize centroids: f32 → u8 [0, 255] (done once, amortized across columns)
+    // Find global min/max across all centroid values for uniform quantization
+    let mut c_min = f32::MAX;
+    let mut c_max = f32::MIN;
+    for v in &compiled.centroids[..k * dim] {
+        if *v < c_min { c_min = *v; }
+        if *v > c_max { c_max = *v; }
+    }
+    let c_range = (c_max - c_min).max(1e-10);
+    let c_scale = c_range / 255.0;
+
+    let centroids_u8: Vec<u8> = compiled.centroids[..k * dim].iter()
+        .map(|&v| (((v - c_min) / c_range) * 255.0).round().clamp(0.0, 255.0) as u8)
+        .collect();
+
+    // Select VNNI dot function (same tiered dispatch as build_distance_table_vnni)
+    let dot_fn: fn(&[u8], &[i8]) -> i32 = {
+        #[cfg(target_arch = "x86_64")]
+        {
+            if is_x86_feature_detected!("avx512vnni") {
+                |a, b| {
+                    // SAFETY: avx512vnni confirmed
+                    unsafe { ndarray::simd_amx::vnni_dot_u8_i8(a, b) }
+                }
+            } else {
+                ndarray::simd_amx::vnni_dot_u8_i8_scalar
+            }
+        }
+        #[cfg(not(target_arch = "x86_64"))]
+        { ndarray::simd_amx::vnni_dot_u8_i8_scalar }
+    };
+
+    for j in 0..n {
+        // Extract input column j and quantize to i8 [-128, 127]
+        let mut col_f32 = vec![0.0f32; dim];
+        for d in 0..dim {
+            col_f32[d] = _rhs[[d, j]].elem::<f64>() as f32;
+        }
+        let mut x_min = f32::MAX;
+        let mut x_max = f32::MIN;
+        for &v in &col_f32 {
+            if v < x_min { x_min = v; }
+            if v > x_max { x_max = v; }
+        }
+        let x_range = (x_max - x_min).max(1e-10);
+        let x_scale = x_range / 255.0;
+
+        let col_i8: Vec<i8> = col_f32.iter()
+            .map(|&v| (((v - x_min) / x_range) * 255.0).round().clamp(0.0, 255.0) as u8 as i8)
+            .collect();
+
+        // VNNI dot: 256 centroid dots at 64 MACs/instruction
+        let mut centroid_out = vec![0.0f64; k];
+        for c in 0..k {
+            let c_row = &centroids_u8[c * dim..(c + 1) * dim];
+            let raw_dot = dot_fn(c_row, &col_i8);
+
+            // Dequantize: raw_dot was computed on quantized values.
+            // Approximate: result ≈ c_scale × x_scale × raw_dot + bias_correction
+            // The bias from zero-point offsets: sum(c_u8) × x_zero + sum(x_u8) × c_zero + ...
+            // For speed: use the linear approximation (sufficient for inference)
+            centroid_out[c] = raw_dot as f64 * c_scale as f64 * x_scale as f64;
+        }
+
+        // Broadcast via palette assignment
+        for i in 0..m {
+            let c_idx = compiled.assignments[i] as usize;
+            out[[i, j]] = centroid_out[c_idx.min(k - 1)].elem();
+        }
+    }
+
+    true
+}
+
+/// Count of registered compiled linear layers.
+#[cfg(feature = "std")]
+pub fn compiled_linear_count() -> usize {
+    LINEAR_CACHE.read().unwrap().len()
+}
+
+/// Clear all compiled linear layers.
+#[cfg(feature = "std")]
+pub fn clear_compiled_linear_cache() {
+    LINEAR_CACHE.write().unwrap().clear();
+}
+//
 // For quantized u8×i8 matmul (codebook distance table build):
 //   Input A: [m, k] u8 (codebook rows, quantized)
 //   Input B: [k, n] i8 (codebook cols, quantized)
@@ -355,6 +522,13 @@ pub(crate) fn matmul<E: NdArrayElement>(
                     .get()
                     .slice_mut(s!(out_batch, .., ..));
 
+                // Try compiled linear (centroid matmul, O(256) per column).
+                // Falls through to BLAS if no compiled layer matches.
+                #[cfg(feature = "std")]
+                if try_compiled_linear(&lhs_slice, &rhs_slice, &mut out_slice, m, k, n) {
+                    return;
+                }
+
                 // Try compiled attention table (O(1) per element).
                 // Falls through to BLAS if no table is registered for d_head=k.
                 #[cfg(feature = "std")]
diff --git a/crates/burn/upstream b/crates/burn/upstream
index ed72d2b1..76299209 160000
--- a/crates/burn/upstream
+++ b/crates/burn/upstream
@@ -1 +1 @@
-Subproject commit ed72d2b125a364aff18aed2a53396c128e01cb42
+Subproject commit 76299209e63b03236b5bb9d51ae45a22404cacaf