diff --git a/koboldcpp.dll b/koboldcpp.dll
index e8ca6e513..b682d241c 100644
Binary files a/koboldcpp.dll and b/koboldcpp.dll differ
diff --git a/koboldcpp_blas.dll b/koboldcpp_blas.dll
index 15c20c12d..0c1cfa481 100644
Binary files a/koboldcpp_blas.dll and b/koboldcpp_blas.dll differ
diff --git a/otherarch/gpt2_v1.cpp b/otherarch/gpt2_v1.cpp
index fd11d17a2..f40e1feae 100644
--- a/otherarch/gpt2_v1.cpp
+++ b/otherarch/gpt2_v1.cpp
@@ -85,6 +85,8 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
 
     auto & ctx = model.ctx;
 
+    auto memory_type = GGML_V1_TYPE_F16;
+
     size_t ctx_size = 0;
 
     {
@@ -119,8 +121,8 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
         ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_size(wtype));         // c_mlp_proj_w
         ctx_size += n_layer*(         n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(memory_type); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(memory_type); // memory_v
 
         ctx_size += (6 + 12*n_layer)*256; // object overhead
 
@@ -218,8 +220,8 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
         const int n_mem      = n_layer*n_ctx;
         const int n_elements = n_embd*n_mem;
 
-        model.memory_k = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_elements);
-        model.memory_v = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_elements);
+        model.memory_k = ggml_v1_new_tensor_1d(ctx, memory_type, n_elements);
+        model.memory_v = ggml_v1_new_tensor_1d(ctx, memory_type, n_elements);
 
         const size_t memory_size = ggml_v1_nbytes(model.memory_k) + ggml_v1_nbytes(model.memory_v);
 
diff --git a/otherarch/gptj_v1.cpp b/otherarch/gptj_v1.cpp
index a51ee21ab..e88f5917a 100644
--- a/otherarch/gptj_v1.cpp
+++ b/otherarch/gptj_v1.cpp
@@ -103,6 +103,8 @@ ModelLoadResult legacy_gptj_model_load(const std::string & fname, gptj_model_v1
 
     auto & ctx = model.ctx;
 
+    auto memory_type = GGML_V1_TYPE_F16;
+
     size_t ctx_size = 0;
 
     {
@@ -136,8 +138,8 @@ ModelLoadResult legacy_gptj_model_load(const std::string & fname, gptj_model_v1
         ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_sizef(wtype));         // c_mlp_proj_w_trans
         ctx_size += n_layer*(         n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32)); // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_sizef(GGML_V1_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_sizef(memory_type); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_sizef(memory_type); // memory_v
 
         ctx_size += (5 + 10*n_layer)*256; // object overhead
 
@@ -240,8 +242,8 @@ ModelLoadResult legacy_gptj_model_load(const std::string & fname, gptj_model_v1
         const int n_mem      = n_layer*n_ctx;
         const int n_elements = n_embd*n_mem;
 
-        model.memory_k = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_elements);
-        model.memory_v = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_elements);
+        model.memory_k = ggml_v1_new_tensor_1d(ctx, memory_type, n_elements);
+        model.memory_v = ggml_v1_new_tensor_1d(ctx, memory_type, n_elements);
 
         const size_t memory_size = ggml_v1_nbytes(model.memory_k) + ggml_v1_nbytes(model.memory_v);
 
diff --git a/otherarch/gptj_v2.cpp b/otherarch/gptj_v2.cpp
index cdb1ae4b7..511824c16 100644
--- a/otherarch/gptj_v2.cpp
+++ b/otherarch/gptj_v2.cpp
@@ -103,6 +103,8 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
 
     auto & ctx = model.ctx;
 
+    auto memory_type = GGML_TYPE_F16;
+
     size_t ctx_size = 0;
 
     {
@@ -136,8 +138,8 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
         ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
         ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_v
 
         ctx_size += (5 + 10*n_layer)*256; // object overhead
 
@@ -234,8 +236,8 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
         const int n_mem      = n_layer*n_ctx;
         const int n_elements = n_embd*n_mem;
 
-        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model.memory_k = ggml_new_tensor_1d(ctx, memory_type, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, memory_type, n_elements);
 
         const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);