diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 5071556..6b659e2 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -1463,12 +1463,12 @@ bool llama_model_loader::load_all_data(
     std::vector<void *> host_ptrs;
     size_t buffer_idx = 0; // buffer to use for async loads
     ggml_backend_t upload_backend = [&](const char * func) -> ggml_backend_t {
-        if (use_mmap || check_tensors) {
+        if (check_tensors) {
             return nullptr;
         }
         // When not using mmaped io use async uploads from pinned memory to GPU memory.
         // First determine if the backend supports the necessary features for async uploads.
-        auto * buf = bufs.count(0) ? bufs.at(0) : nullptr;
+        auto * buf = bufs.count(0) ? bufs.at(0).buffer : nullptr;
         if (!buf) {
             LLAMA_LOG_DEBUG("%s: no buffer found for async uploads\n", func);
             return nullptr;
@@ -1539,7 +1539,7 @@ bool llama_model_loader::load_all_data(
     if (upload_backend) {
         LLAMA_LOG_DEBUG("%s: using async uploads for device %s, buffer type %s, backend %s\n", __func__,
             ggml_backend_dev_name(ggml_backend_get_device(upload_backend)),
-            ggml_backend_buft_name(ggml_backend_buffer_get_type(bufs.at(0))),
+            ggml_backend_buft_name(ggml_backend_buffer_get_type(bufs.at(0).buffer)),
             ggml_backend_name(upload_backend));
     }
 
@@ -1558,11 +1558,15 @@ bool llama_model_loader::load_all_data(
 
         size_t n_size = ggml_nbytes(cur);
 
-        if (use_mmap) {
+        const auto buf_it = bufs.find(weight->idx);
+        const bool tensor_uses_mmap =
+            use_mmap && buf_it != bufs.end() && buf_it->second.is_mmap;
+
+        if (tensor_uses_mmap) {
             const auto & mapping = mappings.at(weight->idx);
             ggml_backend_buffer_t buf_mmap = nullptr;
-            if (bufs.count(weight->idx)) {
-                buf_mmap = bufs.at(weight->idx);
+            if (buf_it != bufs.end()) {
+                buf_mmap = buf_it->second.buffer;
             }
             uint8_t * data = (uint8_t *) mapping->addr() + weight->offs;
 
diff --git a/src/llama-model-loader.h b/src/llama-model-loader.h
index e9fe359..4b1215d 100644
--- a/src/llama-model-loader.h
+++ b/src/llama-model-loader.h
@@ -15,7 +15,12 @@
 #include <stdexcept>
 #include <unordered_map>
 
-using llama_buf_map = std::unordered_map<uint32_t, ggml_backend_buffer_t>;
+struct llama_buf_info {
+    ggml_backend_buffer_t buffer;
+    bool is_mmap;
+};
+
+using llama_buf_map = std::unordered_map<uint32_t, llama_buf_info>;
 
 // lists of buffer types used for each layer
 using buft_list_t = std::vector<std::pair<ggml_backend_dev_t, ggml_backend_buffer_type_t>>;
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index a71b1aa..388e9dc 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -1664,7 +1664,10 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         }
     }
 
-    ml.init_mappings(true, use_mlock ? &pimpl->mlock_mmaps : nullptr);
+    // Do not prefetch every byte of every shard. Qwen3.8 Flash Next can keep its
+    // sparse PLE tensor mmap-backed on CPU while uploading dense weights to GPU;
+    // whole-file prefetch defeats that arrangement and thrashes memory on UMA.
+    ml.init_mappings(false, use_mlock ? &pimpl->mlock_mmaps : nullptr);
     pimpl->mappings.reserve(ml.mappings.size());
 
     // create the backend buffers
@@ -1720,7 +1723,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
                     throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
                 }
                 bufs.emplace_back(buf);
-                buf_map.emplace(idx, buf);
+                buf_map.emplace(idx, llama_buf_info { buf, true });
             }
         } else {
             ggml_backend_buffer_t buf;
@@ -1743,7 +1746,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
             }
             bufs.emplace_back(buf);
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
-                buf_map.emplace(idx, buf);
+                buf_map.emplace(idx, llama_buf_info { buf, false });
             }
         }