diff --git a/expose.h b/expose.h
index 855779a23..623fb2134 100644
--- a/expose.h
+++ b/expose.h
@@ -41,7 +41,7 @@ struct load_model_inputs
     const char * lora_filename = nullptr;
     const char * lora_base = nullptr;
     const char * draftmodel_filename = nullptr;
-    const int draft_amount = 16;
+    const int draft_amount = 12;
     const char * mmproj_filename = nullptr;
     const bool use_mmap = false;
     const bool use_mlock = false;
diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp
index 9000fed56..9fab4253f 100644
--- a/gpttype_adapter.cpp
+++ b/gpttype_adapter.cpp
@@ -53,7 +53,7 @@ std::string lora_filename = "";
 std::string lora_base = "";
 std::string mmproj_filename = "";
 std::string draftmodel_filename = "";
-int speculative_chunk_amt = 16; //do it in chunks of this many tokens
+int speculative_chunk_amt = 12; //do it in chunks of this many tokens
 bool generation_finished;
 float last_process_time = 0;
 float last_eval_time = 0;
@@ -680,7 +680,10 @@ static speculative_draft_result speculative_decoding_eval_chunk(llama_context *
         ++draft_npast;
     }
     //now that we have our drafted tokens, we form a batch and PP it
-    kcpp_embd_batch batch2 = kcpp_embd_batch(drafted_ids, actual_npast, true);
+
+    std::vector<int> real_embd = drafted_ids;
+    real_embd.pop_back();
+    kcpp_embd_batch batch2 = kcpp_embd_batch(real_embd, actual_npast, true);
     auto draftok = (llama_decode(main_ctx, batch2.batch)==0); //actual eval for big model
     if(!draftok)
     {
diff --git a/koboldcpp.py b/koboldcpp.py
index 839a6c1d0..ef48f8c87 100644
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -2764,7 +2764,7 @@ def show_gui():
     preloadstory_var = ctk.StringVar()
     mmproj_var = ctk.StringVar()
     draftmodel_var = ctk.StringVar()
-    draftamount_var = ctk.StringVar(value="16")
+    draftamount_var = ctk.StringVar(value="12")
     nomodel = ctk.IntVar(value=0)
 
     port_var = ctk.StringVar(value=defaultport)
@@ -3508,7 +3508,7 @@ def show_gui():
             pass
         args.mmproj = None if mmproj_var.get() == "" else mmproj_var.get()
         args.draftmodel = None if draftmodel_var.get() == "" else draftmodel_var.get()
-        args.draftamount = int(draftamount_var.get()) if draftamount_var.get()!="" else 16
+        args.draftamount = int(draftamount_var.get()) if draftamount_var.get()!="" else 12
 
         args.ssl = None if (ssl_cert_var.get() == "" or ssl_key_var.get() == "") else ([ssl_cert_var.get(), ssl_key_var.get()])
         args.password = None if (password_var.get() == "") else (password_var.get())
@@ -4951,7 +4951,7 @@ if __name__ == '__main__':
     advparser.add_argument("--nocertify", help="Allows insecure SSL connections. Use this if you have cert errors and need to bypass certificate restrictions.", action='store_true')
     advparser.add_argument("--mmproj", help="Select a multimodal projector file for vision models like LLaVA.", default="")
     advparser.add_argument("--draftmodel", help="Load a small draft model for speculative decoding. It will be fully offloaded. Vocab must match the main model.", default="")
-    advparser.add_argument("--draftamount", metavar=('[tokens]'), help="How many tokens to draft per chunk before verifying results", type=int, default=16)
+    advparser.add_argument("--draftamount", metavar=('[tokens]'), help="How many tokens to draft per chunk before verifying results", type=int, default=12)
     advparser.add_argument("--password", help="Enter a password required to use this instance. This key will be required for all text endpoints. Image endpoints are not secured.", default=None)
     advparser.add_argument("--ignoremissing", help="Ignores all missing non-essential files, just skipping them instead.", action='store_true')
     advparser.add_argument("--chatcompletionsadapter", help="Select an optional ChatCompletions Adapter JSON file to force custom instruct tags.", default="")