diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index d43747543d..390d3543d7 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -161,7 +161,8 @@ struct llama_hparams {
     // the size of the sliding window (0 - no SWA)
     uint32_t n_swa = 0;
 
-    // deepseek4 vision: when set to non-causal attention, SWA should not be applied (case: input is multimodal)
+    // deepseek4 vision: when decoding non-causally (multimodal input), SWA is not applied between tokens of the current ubatch (the image span); older tokens are still window-clipped
+    // for other models (like gemma 3, gemma 4): SWA is always applied to match transformers implementation
     bool swa_full_non_causal = false;
 
     // if is_swa_impl[il] == 1, then layer il is SWA