# Expose the size of the loaded weights, for about:memory. Upstreamable to https://github.com/mudler/parakeet.cpp
diff --git a/include/parakeet_capi.h b/include/parakeet_capi.h
--- a/include/parakeet_capi.h
+++ b/include/parakeet_capi.h
@@ -1,6 +1,8 @@
 #ifndef PARAKEET_CAPI_H
 #define PARAKEET_CAPI_H
 
+#include <stddef.h>
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -54,6 +56,9 @@
 // Free a context obtained from parakeet_capi_load. Safe on NULL.
 void parakeet_capi_free(parakeet_ctx* ctx);
 
+// Bytes the context's weights occupy, for memory reporting. 0 on NULL.
+size_t parakeet_capi_weights_bytes(const parakeet_ctx* ctx);
+
 // Transcribe a WAV file. `decoder` selects the head:
 //   0 = default (by arch: transducer for tdt/rnnt/hybrid, CTC for ctc),
 //   1 = ctc (force CTC head),
diff --git a/src/model.hpp b/src/model.hpp
--- a/src/model.hpp
+++ b/src/model.hpp
@@ -93,6 +93,8 @@
     // The underlying loaded GGUF. Exposed so the streaming C-API can build a
     // pk::StreamingSession (and a MelFrontend) over the same load-once model.
     const ModelLoader& loader() const { return loader_; }
+    // Bytes the loaded weights occupy, for memory reporting.
+    size_t weights_bytes() const { return loader_.weights_bytes(); }
 
     // Non-copyable (owns the GGUF mapping).
     Model(const Model&) = delete;
diff --git a/src/model_loader.hpp b/src/model_loader.hpp
--- a/src/model_loader.hpp
+++ b/src/model_loader.hpp
@@ -95,6 +95,8 @@
     // the same CPU backend the compute path uses. Returns false on failure.
     bool realize_weights(ggml_backend_t backend);
     bool weights_realized() const { return weights_buf_ != nullptr; }
+    // Bytes the loaded weights occupy, for memory reporting.
+    size_t weights_bytes() const;
 private:
     // Parse metadata + tensors from an already-opened gguf_. Shared by load()
     // and load_fd().
diff --git a/src/model_loader.cpp b/src/model_loader.cpp
--- a/src/model_loader.cpp
+++ b/src/model_loader.cpp
@@ -71,6 +71,12 @@
     if(device_ctx_) ggml_free(device_ctx_);
     if(gguf_) gguf_free(gguf_); if(ctx_) ggml_free(ctx_);
 }
+// The backend buffer once realized, else ctx_'s mem_buffer.
+size_t ModelLoader::weights_bytes() const {
+    if (weights_buf_) return ggml_backend_buffer_get_size(weights_buf_);
+    return ctx_ ? ggml_get_mem_size(ctx_) : 0;
+}
+
 bool ModelLoader::realize_weights(ggml_backend_t backend){
     if(weights_buf_) return true;                       // idempotent
     if(!backend || !ctx_){ PK_LOG("realize_weights: null backend/ctx"); return false; }
diff --git a/src/parakeet_capi.cpp b/src/parakeet_capi.cpp
--- a/src/parakeet_capi.cpp
+++ b/src/parakeet_capi.cpp
@@ -225,6 +225,11 @@
     }
 }
 
+extern "C" size_t parakeet_capi_weights_bytes(const parakeet_ctx* ctx) {
+    if (!ctx || !ctx->model) return 0;
+    return ctx->model->weights_bytes();
+}
+
 // Firefox-local: load from an already-open fd (sandboxed host).
 extern "C" parakeet_ctx* parakeet_capi_load_fd(int fd) {
     if (fd < 0) return nullptr;