diff --git a/app/server/README.md b/app/server/README.md
index e9e49d862..790a49c62 100644
--- a/app/server/README.md
+++ b/app/server/README.md
@@ -397,3 +397,42 @@ curl http://127.0.0.1:8080/v1/tasks/run \
     }
   }'
 ```
+
+### `POST /v1/tasks/unload_models`
+
+Unload specific models from memory to free resources (e.g. VRAM on GPU backends). Subsequent requests to an unloaded model will trigger a transparent reload. The server waits for any in-flight inference on each target model to complete before unloading it.
+
+The request body must contain a `model_ids` array of strings. Unknown ids are reported in the response rather than causing an error. Models that are not yet loaded (e.g. lazy-loaded models that have not been requested yet) are skipped silently.
+
+```bash
+curl http://127.0.0.1:8080/v1/tasks/unload_models \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model_ids": ["pocket-tts", "qwen3-asr"]
+  }'
+```
+
+Response:
+
+```json
+{
+  "unloaded": ["pocket-tts", "qwen3-asr"],
+  "not_found": []
+}
+```
+
+### `POST /v1/tasks/unload_all_models`
+
+Unload all currently loaded models from memory. No request body is required. As with the selective endpoint, subsequent requests will reload models transparently.
+
+```bash
+curl -X POST http://127.0.0.1:8080/v1/tasks/unload_all_models
+```
+
+Response:
+
+```json
+{
+  "unloaded": ["pocket-tts", "qwen3-asr"]
+}
+```
diff --git a/app/server/runtime.cpp b/app/server/runtime.cpp
index 46ce9755a..3d1250d21 100644
--- a/app/server/runtime.cpp
+++ b/app/server/runtime.cpp
@@ -686,6 +686,12 @@ HttpResponse ServerState::handle(const HttpRequest & request) {
     else if (request.method == "POST" && request.path == "/v1/tasks/stream") {
         response = handle_generic_stream(request.body);
     }
+    else if (request.method == "POST" && request.path == "/v1/tasks/unload_models") {
+        response = handle_unload_models(request.body);
+    }
+    else if (request.method == "POST" && request.path == "/v1/tasks/unload_all_models") {
+        response = handle_unload_all_models();
+    }
     else {
         response = error_response(404, "unknown endpoint: " + request.path, "not_found");
     }
@@ -1544,4 +1550,78 @@ std::string ServerState::get_allowed_origin(const HttpRequest & request) const {
     return "";
 }
 
+void ServerState::LoadedModel::unload() {
+	offline = nullptr;
+    streaming = nullptr;
+    session.reset();
+    model.reset();
+}
+
+HttpResponse ServerState::handle_unload_models(const std::string & body_text) {
+    const auto body = engine::io::json::parse(body_text);
+    const auto * ids = body.find("model_ids");
+    if (ids == nullptr || !ids->is_array()) {
+        return error_response(400, "request requires a 'model_ids' string array", "invalid_request_error");
+    }
+
+    std::vector<std::string> unloaded;
+    std::vector<std::string> not_found;
+
+    for (const auto & id_val : ids->as_array()) {
+        if (!id_val.is_string()) {
+            return error_response(400, "each element of 'model_ids' must be a string", "invalid_request_error");
+        }
+        const std::string id = id_val.as_string();
+        const auto it = model_index_.find(id);
+        if (it == model_index_.end()) {
+            not_found.push_back(id);
+            continue;
+        }
+        LoadedModel & model = *models_.at(it->second);
+        // Only unload if the model is currently loaded in memory. Acquire the busy
+        // lock for the duration of the unload so no inference starts mid-operation.
+        if (model.session != nullptr) {
+            [[maybe_unused]] BusyGuard::Lock lock = model.busy.acquire(0, model.config.id);
+            model.unload();
+            unloaded.push_back(id);
+        }
+    }
+
+    std::ostringstream out;
+    out << "{\"unloaded\":[";
+    for (size_t i = 0; i < unloaded.size(); ++i) {
+        if (i != 0) { out << ","; }
+        out << json_quote(unloaded[i]);
+    }
+    out << "],\"not_found\":";
+    out << "[";
+    for (size_t i = 0; i < not_found.size(); ++i) {
+        if (i != 0) { out << ","; }
+        out << json_quote(not_found[i]);
+    }
+    out << "]}";
+    return json_response(out.str());
+}
+
+HttpResponse ServerState::handle_unload_all_models() {
+    std::vector<std::string> unloaded;
+
+    for (auto & model : models_) {
+        if (model->session != nullptr) {
+            [[maybe_unused]] BusyGuard::Lock lock = model->busy.acquire(0, model->config.id);
+            model->unload();
+            unloaded.push_back(model->config.id);
+        }
+    }
+
+    std::ostringstream out;
+    out << "{\"unloaded\":[";
+    for (size_t i = 0; i < unloaded.size(); ++i) {
+        if (i != 0) { out << ","; }
+        out << json_quote(unloaded[i]);
+    }
+    out << "]}";
+    return json_response(out.str());
+}
+
 }  // namespace minitts::server
diff --git a/app/server/runtime.h b/app/server/runtime.h
index 298dfe369..44b965a0a 100644
--- a/app/server/runtime.h
+++ b/app/server/runtime.h
@@ -52,6 +52,10 @@ class ServerState final : public IHttpHandler {
         // Serializes runs on this model and bounds how long a caller waits for its
         // turn; see BusyGuard.
         BusyGuard busy;
+
+        // Release the loaded model and session from memory (frees VRAM on GPU backends).
+        // The next request will trigger a reload via ensure_model_loaded_locked().
+        void unload();
     };
 
     // Acquire the model's run guard. `request_timeout_ms` is the caller-supplied
@@ -125,6 +129,8 @@ class ServerState final : public IHttpHandler {
     HttpResponse handle_generic_run(const std::string & body_text);
     HttpResponse handle_generic_stream(const std::string & body_text);
     HttpResponse handle_voices(const HttpRequest & request) const;
+    HttpResponse handle_unload_models(const std::string & body_text);
+    HttpResponse handle_unload_all_models();
     std::string models_json() const;
     std::string get_allowed_origin(const HttpRequest & request) const;