assistant: unload the resident model from the settings page

The Load button reads Unload on the loaded model and posts to the
router's /models/unload. An error from the router lands in modelError,
the same place a failed load reports, and the list is re-read either way.
The model comes back on the next request for it or when the canvas opens.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

2 files changed, 18 insertions, 4 deletions

diff --git a/desktop/modules/assistant/AssistantModule.qml b/desktop/modules/assistant/AssistantModule.qml
index a05556d..9e8adbc 100644
--- a/desktop/modules/assistant/AssistantModule.qml
+++ b/desktop/modules/assistant/AssistantModule.qml
@@ -362,6 +362,18 @@ Module {
             else { poll.left = 300; poll.start(); }
         });
     }
+    // Frees the VRAM for every client; the next request for the model, or
+    // the canvas opening, loads it again.
+    function unloadModel(id) {
+        am.modelError = "";
+        am.run(["curl", "-s", "-m", "30", "-X", "POST", "-H", "Content-Type: application/json",
+                "-d", JSON.stringify({ model: id }), am.routerRoot + "/models/unload"], null, (code, out) => {
+            let err = code === 0 ? "" : "the router did not answer";
+            try { const e = JSON.parse(out).error; if (e) err = e.message; } catch (x) {}
+            if (err) am.modelError = `${id}: ${err}`;
+            am.listModels();
+        });
+    }
     // Called when the canvas opens and when its chat's model changes, with
     // the chat's model, else the preferred one: loaded if it is a router
     // model neither loaded nor loading. A cloud id is not in the router's
diff --git a/desktop/modules/assistant/AssistantPage.qml b/desktop/modules/assistant/AssistantPage.qml
index bf35763..b25b33b 100644
--- a/desktop/modules/assistant/AssistantPage.qml
+++ b/desktop/modules/assistant/AssistantPage.qml
@@ -168,7 +168,8 @@ Column {
         wrapMode: Text.Wrap
         text: "Each chat uses the model picked in the canvas header, else the preferred one. "
             + "Opening the canvas loads that model if it is a local one and another is in. "
-            + "Load swaps a model in now, for every client, without changing the preference."
+            + "Load swaps a model in now, for every client, without changing the preference; "
+            + "Unload frees the VRAM until something asks for the model again."
     }
 
     Small {
@@ -218,9 +219,10 @@ Column {
             Button {
                 id: loadBtn
                 anchors { right: parent.right; rightMargin: 6; verticalCenter: parent.verticalCenter }
-                text: "Load"
-                enabled: page.asst.loading === "" && !modelRow.loaded
-                onClicked: page.asst.loadModel(modelRow.modelData.id)
+                text: modelRow.loaded ? "Unload" : "Load"
+                enabled: page.asst.loading === ""
+                onClicked: modelRow.loaded ? page.asst.unloadModel(modelRow.modelData.id)
+                                           : page.asst.loadModel(modelRow.modelData.id)
             }
             Button {
                 anchors { right: loadBtn.left; rightMargin: 6; verticalCenter: parent.verticalCenter }