diff --git a/ui/src/ui/chat-model-ref.test.ts b/ui/src/ui/chat-model-ref.test.ts
index a0ec9d749b8..4e26a913c14 100644
--- a/ui/src/ui/chat-model-ref.test.ts
+++ b/ui/src/ui/chat-model-ref.test.ts
@@ -29,6 +29,12 @@ describe("chat-model-ref helpers", () => {
     });
   });
 
+  it("preserves already-qualified model refs without prepending provider", () => {
+    expect(resolveServerChatModelValue("ollama/qwen3:30b", "openai-codex")).toBe(
+      "ollama/qwen3:30b",
+    );
+  });
+
   it("normalizes raw overrides when the catalog match is unique", () => {
     expect(normalizeChatModelOverrideValue(createChatModelOverride("gpt-5-mini"), catalog)).toBe(
       "openai/gpt-5-mini",
diff --git a/ui/src/ui/chat-model-ref.ts b/ui/src/ui/chat-model-ref.ts
index 7b54ec5e2b0..deb819fa8e3 100644
--- a/ui/src/ui/chat-model-ref.ts
+++ b/ui/src/ui/chat-model-ref.ts
@@ -25,6 +25,11 @@ export function buildQualifiedChatModelValue(model: string, provider?: string |
   if (!trimmedModel) {
     return "";
   }
+  // Preserve already-qualified model refs (provider/model) as-is.
+  // This avoids prepending an unrelated/default provider.
+  if (trimmedModel.includes("/")) {
+    return trimmedModel;
+  }
   const trimmedProvider = provider?.trim();
   return trimmedProvider ? `${trimmedProvider}/${trimmedModel}` : trimmedModel;
 }