diff --git a/app/server/runtime.cpp b/app/server/runtime.cpp
index 03f1fcf5..74d1e380 100644
--- a/app/server/runtime.cpp
+++ b/app/server/runtime.cpp
@@ -1180,9 +1180,26 @@ std::unique_ptr<ServerState::LoadedModel> ServerState::make_model(ServerModelCon
         engine::runtime::parse_run_mode(loaded->config.mode),
     };
     load_voice_presets(*loaded);
+    refresh_model_option_flags(*loaded);
     return loaded;
 }
 
+void ServerState::refresh_model_option_flags(LoadedModel & model) {
+    const auto effective_override = model.config.model_spec_override.has_value()
+        ? model.config.model_spec_override
+        : config_.model_spec_override;
+    // Deliberately uncaught: model_accepts_request_option already returns true for a
+    // model with no contract, swallowing only the missing-contract errors and
+    // rethrowing the rest. Anything that propagates here is therefore a real
+    // misconfiguration (invalid spec, missing override file, family mismatch) and
+    // must fail at registration rather than be assumed away.
+    model.accepts_reference_text = model_accepts_request_option(
+        model.config.family,
+        "reference_text",
+        effective_override,
+        model.config.path);
+}
+
 HttpResponse ServerState::handle_model_load(const std::string & body_text) {
     if (!config_.ui_management) {
         return error_response(403, "dynamic model management is disabled", "forbidden");
@@ -1221,6 +1238,7 @@ HttpResponse ServerState::handle_model_load(const std::string & body_text) {
                 engine::runtime::parse_run_mode(existing->config.mode),
             };
             load_voice_presets(*existing);
+            refresh_model_option_flags(*existing);
         }
         ensure_model_loaded_locked(*existing);
         return json_response(
@@ -1869,15 +1887,10 @@ engine::runtime::TaskRequest ServerState::build_speech_request(const LoadedModel
 
     bool voice_field_is_preset = false;
     const auto * preset = select_voice_preset(model, body, voice_field_is_preset);
-    const auto effective_model_spec_override = model.config.model_spec_override.has_value()
-        ? model.config.model_spec_override
-        : config_.model_spec_override;
-    const bool can_inject_reference_text =
-        model_accepts_request_option(
-            model.config.family,
-            "reference_text",
-            effective_model_spec_override,
-            model.config.path);
+    // Resolved once at registration (refresh_model_option_flags): calling
+    // model_accepts_request_option per request re-reads the model file's embedded
+    // spec on the request thread, which cost ~0.9 s per request for large GGUFs.
+    const bool can_inject_reference_text = model.accepts_reference_text;
 
     engine::runtime::VoiceCondition voice;
     bool has_voice = false;
diff --git a/app/server/runtime.h b/app/server/runtime.h
index d4658885..07c5a52b 100644
--- a/app/server/runtime.h
+++ b/app/server/runtime.h
@@ -65,6 +65,12 @@ class ServerState final : public IHttpHandler {
         mutable std::shared_mutex metadata_mutex;
         std::unordered_map<std::string, RuntimeVoicePreset> voice_presets;
         std::optional<RuntimeVoicePreset> default_voice_preset;
+        // Whether this model's contract accepts the `reference_text` request
+        // option, resolved once at registration (refresh_model_option_flags).
+        // Resolving it per request re-reads the model file's embedded spec on
+        // the request thread, which costs ~0.9 s per request for large GGUFs.
+        // `true` mirrors model_accepts_request_option's no-contract behavior.
+        bool accepts_reference_text = true;
         // Serializes runs on this model and bounds how long a caller waits for its
         // turn; see BusyGuard.
         BusyGuard busy;
@@ -85,6 +91,9 @@ class ServerState final : public IHttpHandler {
 
     void load_models();
     std::unique_ptr<LoadedModel> make_model(ServerModelConfig config);
+    // Recompute the per-model, config-derived request-option flags (currently
+    // accepts_reference_text). Called at registration and on reconfiguration.
+    void refresh_model_option_flags(LoadedModel & model);
     std::filesystem::path resolve_ui_model_path(const std::filesystem::path & path) const;
     HttpResponse handle_model_load(const std::string & body_text);
     HttpResponse handle_model_unload(const std::string & body_text);