diff --git a/crates/spark-model/src/weight_loader/qwen4_exp/probe.rs b/crates/spark-model/src/weight_loader/qwen4_exp/probe.rs
index 4970939d4..3176cb0e9 100644
--- a/crates/spark-model/src/weight_loader/qwen4_exp/probe.rs
+++ b/crates/spark-model/src/weight_loader/qwen4_exp/probe.rs
@@ -29,9 +29,12 @@ pub struct NamespaceReport {
     pub stray_layer_norms: usize,
 }
 
-/// A projection is present in either the dense/NVFP4 layout (`.weight`) or the EXL3 one (`.trellis`).
+/// A projection is present in the dense/NVFP4 layout (`.weight`), the compressed-tensors
+/// one (`.weight_packed`), or the EXL3 one (`.trellis`).
 fn has_linear(store: &WeightStore, stem: &str) -> bool {
-    store.contains(&format!("{stem}.weight")) || store.contains(&format!("{stem}.trellis"))
+    store.contains(&format!("{stem}.weight"))
+        || store.contains(&format!("{stem}.weight_packed"))
+        || store.contains(&format!("{stem}.trellis"))
 }
 
 /// The n-gram tables are deferred (served from NVMe), which `contains` does not see.
diff --git a/crates/spark-model/src/weight_loader/qwen4_exp/probe_tests.rs b/crates/spark-model/src/weight_loader/qwen4_exp/probe_tests.rs
index 2e13117a9..4d4cf2626 100644
--- a/crates/spark-model/src/weight_loader/qwen4_exp/probe_tests.rs
+++ b/crates/spark-model/src/weight_loader/qwen4_exp/probe_tests.rs
@@ -1,7 +1,8 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 
 //! `audit_namespace` must see every projection of an EXL3 checkpoint (`.trellis`
-//! instead of `.weight`) and the deferred n-gram tables (never in the store map).
+//! instead of `.weight`), of a compressed-tensors one (`.weight_packed`), and the
+//! deferred n-gram tables (never in the store map).
 
 use super::*;
 use spark_runtime::gpu::DevicePtr;
@@ -81,6 +82,44 @@ fn exl3_layout_is_counted() {
     assert!(r.ensure_loadable().is_ok());
 }
 
+#[test]
+fn compressed_tensors_layout_is_counted() {
+    // A mixed-precision compressed-tensors checkpoint: the routed experts are NVFP4 and
+    // carry `.weight_packed`, while attention and lm_head are FP8 with a plain `.weight`.
+    // Probing only `.weight` therefore counts every layer's attention but no experts,
+    // and the load is refused with "0 of N layers have routed experts".
+    let c = config();
+    let l0 = c.layer_prefix(0);
+    let l1 = c.layer_prefix(1);
+    let names = [
+        n(format!("{l0}.linear_attn.in_proj_qkv.weight")),
+        n(format!("{l1}.self_attn.q_proj.weight")),
+        n(format!("{l1}.self_attn.indexer.index_qk_proj.weight")),
+        n(format!("{l0}.mlp.experts.0.gate_proj.weight_packed")),
+        n(format!("{l1}.mlp.experts.0.gate_proj.weight_packed")),
+        n(format!("{l0}.attn_hyper_connection.hc_norm.weight")),
+        n(format!("{l1}.attn_hyper_connection.hc_norm.weight")),
+        n(format!("{l0}.mlp_hyper_connection.hc_norm.weight")),
+        n(format!("{l1}.mlp_hyper_connection.hc_norm.weight")),
+        n("model.language_model.embed_tokens.weight".into()),
+        n("lm_head.weight".into()),
+    ];
+    // The n-gram table stays BF16 in this layout and is deferred as numbered shards.
+    let base = format!("{l1}.ple.ple_embedding.ngram_embedding");
+    let deferred = [
+        n(format!("{base}.shard_0.weight")),
+        n(format!("{base}.shard_1.weight")),
+    ];
+    let r = audit_namespace(&store(&names, &deferred), &c);
+    assert_eq!(r.gdn_layers, 1);
+    assert_eq!(r.attn_layers, 1);
+    assert_eq!(r.indexer_tensors, 1);
+    assert_eq!(r.expert_tensors, 2);
+    assert!(r.has_lm_head);
+    assert_eq!(r.ple_shards, 2);
+    assert!(r.ensure_loadable().is_ok());
+}
+
 #[test]
 fn deferred_shards_are_counted() {
     let c = config();