diff --git a/common/sampling.cpp b/common/sampling.cpp
index 1317024c2..ca4e6bbb7 100644
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -187,13 +187,16 @@ llama_token llama_sampling_sample(
         } else {
             // temperature sampling
             size_t min_keep = std::max(1, params.n_probs);
+            float output_temp = 1.0; // WIP, this should be defined separately as its own parameter.
 
+            llama_sample_temp     (ctx_main, &cur_p, temp); // "Initial temp", this is how it's traditionally implemented
             llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep);
             llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep);
             llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep);
             llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep);
             llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep);
-            llama_sample_temp     (ctx_main, &cur_p, temp);
+            llama_sample_temp     (ctx_main, &cur_p, temp_last); // "Output temp", this is how llama.cpp chose to implement it
+
 
             id = llama_sample_token(ctx_main, &cur_p);