diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 637f8d216475..4d4913312101 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -92,6 +92,10 @@ struct server_slot {
     std::vector<int32_t> spec_i_batch;
     server_prompt_checkpoint spec_ckpt;
     common_speculative_ptr spec;
+    bool spec_runtime_disabled = false;
+    int32_t spec_verify_batches = 0;
+    int32_t spec_verify_draft_total = 0;
+    int32_t spec_verify_draft_accepted = 0;
 
     // TODO: move members that belong to the task (such as `generated_text`, `has_new_line`) to task_results_state
     //       see https://github.com/ggml-org/llama.cpp/pull/18283#issuecomment-3710175837
@@ -218,6 +222,10 @@ struct server_slot {
         // clear speculative decoding stats
         n_draft_total = 0;
         n_draft_accepted = 0;
+        spec_runtime_disabled = false;
+        spec_verify_batches = 0;
+        spec_verify_draft_total = 0;
+        spec_verify_draft_accepted = 0;
 
         task_prev = std::move(task);
         task.reset();
@@ -310,6 +318,10 @@ struct server_slot {
             return 0;
         }
 
+        if (spec_runtime_disabled && spec_draft.empty()) {
+            return 0;
+        }
+
         const int n_draft_min = common_speculative_n_min(spec.get(), task->params.speculative);
 
         // determine the max draft that fits the current slot state
@@ -2999,6 +3011,24 @@ struct server_context_impl {
 
                     GGML_ASSERT(accepted.size() >= 1);
 
+                    const int32_t n_draft_accepted_actual = std::max<int32_t>(0,
+                            std::min<int32_t>((int32_t) n_draft, (int32_t) accepted.size() - 1));
+                    slot.spec_verify_batches += 1;
+                    slot.spec_verify_draft_total += (int32_t) n_draft;
+                    slot.spec_verify_draft_accepted += n_draft_accepted_actual;
+                    if (slot.task->params.speculative.type == COMMON_SPECULATIVE_TYPE_MTP &&
+                            !slot.spec_runtime_disabled &&
+                            slot.spec_verify_batches >= 2 &&
+                            slot.spec_verify_draft_total >= 16 &&
+                            4 * slot.spec_verify_draft_accepted < 3 * slot.spec_verify_draft_total) {
+                        slot.spec_runtime_disabled = true;
+                        SLT_WRN(slot,
+                                "disabling MTP speculation for this request: actual draft acceptance %d/%d after %d verify batches\n",
+                                slot.spec_verify_draft_accepted,
+                                slot.spec_verify_draft_total,
+                                slot.spec_verify_batches);
+                    }
+
                     // check for partial draft acceptance
                     if (accepted.size() < slot.spec_draft.size() + 1) {
                         if (use_ckpt) {