diff --git a/src/free.c b/src/free.c
index 43b362720..c39745c84 100644
--- a/src/free.c
+++ b/src/free.c
@@ -153,7 +153,14 @@ static inline void mi_free_ex(void* p, size_t* usable) mi_attr_noexcept
   mi_segment_t* const segment = mi_checked_ptr_segment(p,"mi_free");
   if mi_unlikely(segment==NULL) return;
 
+  #if defined(MI_SINGLE_THREADED) && (MI_SINGLE_THREADED)
+  // Single-threaded specialization (ExGen-Malloc, CAL'25): in a program that only ever
+  // uses mimalloc from one thread, every block is thread-local, so we can skip the
+  // thread-id TLS read + atomic load + compare on every free.
+  const bool is_local = true;
+  #else
   const bool is_local = (_mi_prim_thread_id() == mi_atomic_load_relaxed(&segment->thread_id));
+  #endif
   mi_page_t* const page = _mi_segment_page_of(segment, p);
   if (usable!=NULL) { *usable = mi_page_usable_block_size(page); }