3 files changed, 15 insertions, 3 deletions
diff --git a/desktop/modules/ai/AiModule.qml b/desktop/modules/ai/AiModule.qml
index 6df3d6b..d58ebfe 100644
--- a/desktop/modules/ai/AiModule.qml
+++ b/desktop/modules/ai/AiModule.qml
@@ -85,6 +85,13 @@ Module {
         exec(["sh", "-c", 'pkill -TERM -P "$1"; kill -TERM "$1"', "sh", app.detail]);
     }
 
+    // The router frees the model's VRAM; the next request for it loads it
+    // again. Its answer is not read, the next poll shows whether it worked.
+    function unload(model) {
+        exec(["curl", "-s", "-m", "30", "-X", "POST", "-H", "Content-Type: application/json",
+              "-d", JSON.stringify({ model: model }), "http://127.0.0.1:8181/models/unload"]);
+    }
+
     // Detached, so a start outlives a quickshell reload. No exit code comes
     // back; the next poll shows whether it worked.
     function exec(cmd) {
diff --git a/desktop/modules/ai/AiPage.qml b/desktop/modules/ai/AiPage.qml
index e89132a..369c1b5 100644
--- a/desktop/modules/ai/AiPage.qml
+++ b/desktop/modules/ai/AiPage.qml
@@ -12,7 +12,8 @@
 import QtQuick
 import "../.."
 
-// Engines are read-only: stopping either would take its web UI down with it.
+// Engines are never stopped, that would take their web UI down with them;
+// llama's model can be unloaded, which frees the VRAM and leaves the server up.
 Column {
     id: page
 
@@ -42,6 +43,8 @@ Column {
                     return modelData.name === "llama" ? "No model loaded" : "Running";
                 return (modelData.busy ? "Generating" : "Idle") + " · " + modelData.model;
             }
+            action: modelData.name === "llama" && modelData.model !== "" ? "Unload" : ""
+            onTriggered: page.ai.unload(modelData.model)
         }
     }
 
diff --git a/desktop/modules/ai/README.md b/desktop/modules/ai/README.md
index 5aec615..e71159b 100644
--- a/desktop/modules/ai/README.md
+++ b/desktop/modules/ai/README.md
@@ -8,8 +8,10 @@ The tile counts running apps and names any engine that is generating, or says
 ## Engines
 
 llama.cpp (`llama-server`, router mode, port 8181) and stable-diffusion.cpp
-(`sd-server`, started by `~/bin/sd-webui`). Both rows are read-only: each
-engine also serves a web UI, and stopping it would take that down.
+(`sd-server`, started by `~/bin/sd-webui`). Neither is ever stopped: each
+engine also serves a web UI, and stopping it would take that down. llama's
+row offers Unload while a model is loaded, `POST /models/unload` to the
+router, which frees the VRAM and leaves the server up.
 
 llama's state comes from its API: `/models` for the loaded model, then
 `/slots?model=<id>&autoload=false` for whether a slot is processing. The