diff --git a/tools/mtmd/models/internvl.cpp b/tools/mtmd/models/internvl.cpp
index 65d7d5a6b..af8f12cdc 100644
--- a/tools/mtmd/models/internvl.cpp
+++ b/tools/mtmd/models/internvl.cpp
@@ -10,7 +10,7 @@ ggml_cgraph * clip_graph_internvl::build() {
     // add CLS token
     ggml_tensor * cls_repeated = ggml_repeat_4d(ctx0, model.class_embedding,
             model.class_embedding->ne[0], 1, n_batch, 1);
-    inp = ggml_concat(ctx0, inp, cls_repeated, 1);
+    inp = ggml_concat(ctx0, cls_repeated, inp, 1);
 
     // The larger models use a different ViT, which uses RMS norm instead of layer norm
     // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-2869786188
@@ -28,7 +28,7 @@ ggml_cgraph * clip_graph_internvl::build() {
     // remove CLS token
     cur = ggml_view_3d(ctx0, cur,
         n_embd, n_patches, n_batch,
-        cur->nb[1], cur->nb[2], 0);
+        cur->nb[1], cur->nb[2], cur->nb[1]);
     cur = ggml_cont(ctx0, cur);
 
     // pixel shuffle