diff --git a/ggml.c b/ggml.c
index 1d5854960a51b..54365b7ae26c6 100644
--- a/ggml.c
+++ b/ggml.c
@@ -3,6 +3,7 @@
 
 #include "ggml-impl.h"
 #include "ggml-quants.h"
+#include "ggml.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -43,6 +44,10 @@
 
 #if defined(_WIN32)
 
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
 #include <windows.h>
 
 typedef volatile LONG atomic_int;
@@ -430,6 +435,57 @@ int64_t ggml_cycles_per_ms(void) {
 #define ggml_perf_cycles_per_ms() 0
 #endif
 
+//
+// cross-platform UTF-8 file paths
+//
+
+#ifdef _WIN32
+static wchar_t * ggml_mbstowcs(const char * mbs) {
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, NULL, 0);
+    if (!wlen) {
+        errno = EINVAL;
+        return NULL;
+    }
+
+    wchar_t * wbuf = GGML_MALLOC(wlen * sizeof(wchar_t));
+    wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, wbuf, wlen);
+    if (!wlen) {
+        GGML_FREE(wbuf);
+        errno = EINVAL;
+        return NULL;
+    }
+
+    return wbuf;
+}
+#endif
+
+FILE * ggml_fopen(const char * fname, const char * mode) {
+#ifdef _WIN32
+    FILE * file = NULL;
+
+    // convert fname (UTF-8)
+    wchar_t * wfname = ggml_mbstowcs(fname);
+    if (wfname) {
+        // convert mode (ANSI)
+        wchar_t * wmode = GGML_MALLOC(strlen(mode) + 1);
+        wchar_t * wmode_p = wmode;
+        do {
+            *wmode_p++ = (wchar_t)*mode;
+        } while (*mode++);
+
+        // open file
+        file = _wfopen(wfname, wmode);
+
+        GGML_FREE(wfname);
+        GGML_FREE(wmode);
+    }
+
+    return file;
+#else
+    return fopen(fname, mode);
+#endif
+}
+
 //
 // cache line
 //
@@ -18739,7 +18795,7 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
     // write binary data
     {
-        FILE * fout = fopen(fname, "wb");
+        FILE * fout = ggml_fopen(fname, "wb");
 
         if (!fout) {
             fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
@@ -18877,7 +18933,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
     // read file into data
     {
-        FILE * fin = fopen(fname, "rb");
+        FILE * fin = ggml_fopen(fname, "rb");
         if (!fin) {
             fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
             return result;
@@ -19213,7 +19269,7 @@ static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node,
 void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) {
     char color[16];
 
-    FILE * fp = fopen(filename, "w");
+    FILE * fp = ggml_fopen(filename, "w");
     GGML_ASSERT(fp);
 
     fprintf(fp, "digraph G {\n");
@@ -20531,7 +20587,7 @@ struct gguf_context * gguf_init_empty(void) {
 }
 
 struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params) {
-    FILE * file = fopen(fname, "rb");
+    FILE * file = ggml_fopen(fname, "rb");
     if (!file) {
         return NULL;
     }
@@ -21486,7 +21542,7 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf *
 }
 
 void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) {
-    FILE * file = fopen(fname, "wb");
+    FILE * file = ggml_fopen(fname, "wb");
     if (!file) {
         GGML_ASSERT(false && "failed to open file for writing");
     }
diff --git a/ggml.h b/ggml.h
index c937d4a535adb..0a5af72058815 100644
--- a/ggml.h
+++ b/ggml.h
@@ -214,9 +214,10 @@
 #    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#include <stdint.h>
-#include <stddef.h>
 #include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
 
 #define GGML_FILE_MAGIC   0x67676d6c // "ggml"
 #define GGML_FILE_VERSION 1
@@ -708,6 +709,9 @@ extern "C" {
 
     GGML_API void    ggml_print_backtrace(void);
 
+    // accepts a UTF-8 path, even on Windows
+    GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
+
     GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
     GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
diff --git a/llama.cpp b/llama.cpp
index b385ae360fbdb..9614cdb17400f 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -1065,7 +1065,7 @@ struct llama_file {
     size_t size;
 
     llama_file(const char * fname, const char * mode) {
-        fp = std::fopen(fname, mode);
+        fp = ggml_fopen(fname, mode);
         if (fp == NULL) {
             throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
         }
@@ -4006,7 +4006,7 @@ static void llm_load_vocab(
     } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) {
         vocab.linefeed_id = vocab.special_pad_id;
     } else {
-        const std::vector<int> ids = llama_tokenize_internal(vocab, "\u010A", false);
+        const std::vector<int> ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A
         GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         vocab.linefeed_id = ids[0];
     }