feat(app): select an LLM and an image model together (hybrid serve)

The server already supports LUMEN_SERVE=chat|image|hybrid, but the desktop
app stored a single `active_model` and launched with a binary `image_mode`,
so picking an image model overwrote the LLM and hybrid was unreachable.

- Split selection into two independent config slots: `active_model` (LLM) and
  `active_image_model` (diffusion). `#[serde(default)]` keeps old configs
  loadable. Each card toggles its own slot (click again to deselect), so all
  four combinations are reachable: chat, image, hybrid, or none.
- `start_server` derives the serve mode from the two slots and launches the
  server accordingly (both set → hybrid). New `ServeKind` enum replaces the
  `image_mode` bool through `Supervisor::start` / `apply_env`.
- `apply_env` now passes `IMAGE_MODEL_ID` (the server reads this — not
  MODEL_ID — for the diffusion backend). This also fixes a latent bug where
  selecting the bf16 `black-forest-labs/FLUX.2-dev` silently fell back to the
  4-bit default because the id never reached the server. The wired-memory cap
  is skipped whenever diffusion is loaded (image or hybrid), not just image.
- Start button is disabled (with a tooltip) when neither slot is set, matching
  the server-side guard. en/ko strings for the new deselect / no-model states.

chore(release): v0.11.1

Author: rabbitson87

crates/lumen-app/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lumen-app"
-version = "0.11.0"
+version = "0.11.1"
 edition.workspace = true
 rust-version.workspace = true
 license.workspace = true

crates/lumen-app/frontend/package.json

@@ -1,7 +1,7 @@
 {
   "name": "lumen-app-frontend",
   "private": true,
-  "version": "0.11.0",
+  "version": "0.11.1",
   "type": "module",
   "scripts": {
     "predev": "cargo build --manifest-path ../../../Cargo.toml -p lumen-server --release",

crates/lumen-app/frontend/src/App.svelte

@@ -185,6 +185,13 @@
     config?.active_model != null && outdatedModels.has(config.active_model)
   );
 
+  // No model in either slot → nothing to serve. Blocks the Start button (the
+  // server would otherwise reject the launch). Both slots are independent, so
+  // this is only true when chat AND image are both unselected.
+  let noModelSelected = $derived(
+    config?.active_model == null && config?.active_image_model == null
+  );
+
   // True when the configured active model exists on disk but its
   // download is incomplete (truncated shard, missing index, stray
   // .part file). Block the Start button so the server doesn't crash
@@ -231,7 +238,7 @@
   // The active model id, if it is a diffusion image model. Drives both the
   // image-models card's ✓ state and the generation panel's visibility.
   let activeImageModelId = $derived.by(() => {
-    const id = config?.active_model;
+    const id = config?.active_image_model;
     if (!id) return null;
     return imageModels.some((m) => m.id === id) ? id : null;
   });
@@ -995,14 +1002,16 @@
       class={status.state === "running" || status.state === "starting" ? "danger" : "primary"}
       onclick={toggleServer}
       disabled={status.state === "starting" || status.state === "stopping" ||
-        (status.state !== "running" && (activeOutdated || activeBroken))}
+        (status.state !== "running" && (noModelSelected || activeOutdated || activeBroken))}
       title={status.state === "running"
         ? ""
-        : activeBroken
-          ? t("header.title.brokenActive")
-          : activeOutdated
-            ? t("header.title.outdatedActive")
-            : ""}
+        : noModelSelected
+          ? t("header.title.noModel")
+          : activeBroken
+            ? t("header.title.brokenActive")
+            : activeOutdated
+              ? t("header.title.outdatedActive")
+              : ""}
     >
       {status.state === "running" || status.state === "starting" ? t("header.stop") : t("header.start")}
     </button>
@@ -1401,9 +1410,13 @@
           {:else}
             <button
               onclick={() => setActive(m.id)}
-              disabled={isActive || !m.supported}
-              title={!m.supported ? t("models.action.title.unsupported") : ""}
-            >{t("action.use")}</button>
+              disabled={!m.supported}
+              title={!m.supported
+                ? t("models.action.title.unsupported")
+                : isActive
+                  ? t("models.action.title.deselect")
+                  : ""}
+            >{isActive ? t("action.deselect") : t("action.use")}</button>
           {/if}
           <button class="danger" onclick={() => removeModel(m.id)}>{t("action.delete")}</button>
         </div>
@@ -1490,7 +1503,7 @@
     <h2 class={cardH2}>{t("imageModels.title")} <span class="dim">{t("imageModels.titleHint")}</span></h2>
     <div class="flex flex-col gap-2">
       {#each imageModels as im}
-        {@const isActive = config?.active_model === im.id}
+        {@const isActive = config?.active_image_model === im.id}
         {@const fits = !systemInfo || im.min_ram_gb <= systemInfo.ram_gb}
         <div
           class={`image-card relative overflow-hidden flex flex-col gap-1.5 px-3 py-2.5 rounded-md bg-panel-2 border ${
@@ -1508,10 +1521,13 @@
               <div class="dim mono text-[11px]">{im.id} · {im.approx_size_gb}GB · ≥{im.min_ram_gb}GB RAM</div>
             </div>
             <button
-              class="primary"
+              class={isActive ? "" : "primary"}
               onclick={() => setActive(im.id)}
-              disabled={isActive}
-              title={fits ? "" : t("imageModels.action.title.overBudget")}
+              title={isActive
+                ? t("imageModels.action.title.deselect")
+                : fits
+                  ? ""
+                  : t("imageModels.action.title.overBudget")}
             >{isActive ? t("imageModels.active") : t("action.use")}</button>
           </div>
           {#if !fits}

crates/lumen-app/frontend/src/lib/api.ts

@@ -81,6 +81,9 @@ export interface PersistentConfig {
   advanced: AdvancedConfig;
   env_overrides: Record<string, string>;
   active_model: string | null;
+  /** Active diffusion image model id; independent of `active_model` so a chat
+   *  model and an image model can be selected together (→ hybrid serve). */
+  active_image_model: string | null;
   server_binary_path: string | null;
   models_dir: string;
 }

crates/lumen-app/frontend/src/messages/en.ts

@@ -28,6 +28,8 @@ export const en: Record<string, string> = {
     "Active model's download is incomplete. Re-download it first (MODELS card → Re-download).",
   "header.title.outdatedActive":
     "Active model has a newer version on Hub. Update it first (MODELS card → Update).",
+  "header.title.noModel":
+    "Select a chat and/or image model in the MODELS card first.",
 
   // ── Status indicators ───────────────────────────────────────────
   "status.stopped": "stopped",
@@ -40,6 +42,7 @@ export const en: Record<string, string> = {
   "action.download": "Download",
   "action.delete": "Delete",
   "action.use": "Use",
+  "action.deselect": "Deselect",
   "action.update": "Update",
   "action.redownload": "Re-download",
   "action.downloading": "Downloading…",
@@ -213,6 +216,8 @@ export const en: Record<string, string> = {
   "models.action.title.update": "Re-download with the latest Hub weights",
   "models.action.title.unsupported":
     "Not in the server-side supported catalog",
+  "models.action.title.deselect":
+    "Click to deselect — run image-only, or pick a different LLM",
 
   // ── Header memory bar ───────────────────────────────────────────
   "header.memory.title":
@@ -273,7 +278,8 @@ export const en: Record<string, string> = {
   "imageModels.active": "Active",
   "imageModels.overBudget": "Larger than this Mac's RAM — generation may swap or fail.",
   "imageModels.action.title.overBudget": "Exceeds this Mac's RAM budget",
-  "imageModels.startHint": "Press Start to launch the server in image mode, then a generation panel appears below.",
+  "imageModels.action.title.deselect": "Click to deselect this image model",
+  "imageModels.startHint": "Press Start to launch the server in image mode, then a generation panel appears below. Selecting an LLM as well launches hybrid mode (both serve at once).",
 
   // ── Image generation panel ──────────────────────────────────────
   "image.title": "IMAGE GENERATION",

crates/lumen-app/frontend/src/messages/ko.ts

@@ -23,6 +23,7 @@ export const ko: Record<string, string> = {
     "활성 모델의 다운로드가 완료되지 않았습니다. 모델 카드에서 재다운로드를 먼저 진행하세요.",
   "header.title.outdatedActive":
     "활성 모델의 최신 버전이 Hub에 있습니다. 모델 카드에서 업데이트를 먼저 진행하세요.",
+  "header.title.noModel": "모델 카드에서 채팅 또는 이미지 모델을 먼저 선택하세요.",
 
   // ── 상태 표시 ───────────────────────────────────────────────────
   "status.stopped": "중지됨",
@@ -35,6 +36,7 @@ export const ko: Record<string, string> = {
   "action.download": "다운로드",
   "action.delete": "삭제",
   "action.use": "사용",
+  "action.deselect": "선택 해제",
   "action.update": "업데이트",
   "action.redownload": "재다운로드",
   "action.downloading": "다운로드 중…",
@@ -207,6 +209,7 @@ export const ko: Record<string, string> = {
     "누락 또는 잘린 파일을 검증 후 재다운로드",
   "models.action.title.update": "최신 Hub 가중치로 재다운로드",
   "models.action.title.unsupported": "서버 측 지원 카탈로그에 없음",
+  "models.action.title.deselect": "클릭하여 선택 해제 — 이미지 전용으로 실행하거나 다른 LLM 선택",
 
   // ── 헤더 메모리 바 ──────────────────────────────────────────────
   "header.memory.title": "시스템 메모리 — wired + active + compressor",
@@ -266,7 +269,8 @@ export const ko: Record<string, string> = {
   "imageModels.active": "사용 중",
   "imageModels.overBudget": "이 Mac의 RAM보다 큽니다 — 생성 시 스왑되거나 실패할 수 있습니다.",
   "imageModels.action.title.overBudget": "이 Mac의 RAM 예산을 초과합니다",
-  "imageModels.startHint": "시작을 눌러 서버를 이미지 모드로 실행하면 아래에 생성 패널이 나타납니다.",
+  "imageModels.action.title.deselect": "클릭하여 이 이미지 모델 선택 해제",
+  "imageModels.startHint": "시작을 눌러 서버를 이미지 모드로 실행하면 아래에 생성 패널이 나타납니다. LLM도 함께 선택하면 하이브리드 모드(둘 다 동시 서빙)로 실행됩니다.",
 
   // ── 이미지 생성 패널 ────────────────────────────────────────────
   "image.title": "이미지 생성",

crates/lumen-app/src/commands.rs

@@ -172,7 +172,14 @@ pub async fn set_active_model(
         if cat.is_image_model(&model_id) {
             drop(cat);
             let mut g = state.config.lock().await;
-            g.active_model = Some(model_id);
+            // Image models occupy their own slot (independent of the chat model)
+            // so a chat + image pair can be active at once → hybrid serve.
+            // Clicking the already-active image model toggles it back off.
+            g.active_image_model = if g.active_image_model.as_deref() == Some(model_id.as_str()) {
+                None
+            } else {
+                Some(model_id)
+            };
             g.save().map_err(err)?;
             return Ok(g.clone());
         }
@@ -203,7 +210,13 @@ pub async fn set_active_model(
         }
     }
     let mut g = state.config.lock().await;
-    g.active_model = Some(model_id);
+    // Clicking the already-active chat model toggles it off (lets the user run
+    // image-only by deselecting the LLM).
+    g.active_model = if g.active_model.as_deref() == Some(model_id.as_str()) {
+        None
+    } else {
+        Some(model_id)
+    };
     g.save().map_err(err)?;
     Ok(g.clone())
 }
@@ -357,72 +370,88 @@ pub async fn check_model_updates(
 pub async fn start_server(app: AppHandle, state: State<'_, AppState>) -> CmdResult<ServerStatus> {
     let g = state.config.lock().await;
     let mut cfg = g.clone();
-    let active_id = g
-        .active_model
-        .clone()
-        .ok_or_else(|| "no active model — pick one in the MODELS card first".to_string())?;
+    // Two independent slots: a chat/LLM model and an image/diffusion model.
+    // Either or both may be set → chat, image, or hybrid serve.
+    let chat_id = g.active_model.clone();
+    let image_id = g.active_image_model.clone();
     let models_dir = g.models_dir.clone();
     let sup = state.supervisor.clone();
     drop(g);
 
+    if chat_id.is_none() && image_id.is_none() {
+        return Err(
+            "no active model — pick a chat and/or image model in the MODELS card first".to_string(),
+        );
+    }
+
     // Hard-gate the launch when the most-recent revision check flagged the
-    // active model as out of date.  The frontend already greys out the Start
-    // button in that state — this is the belt-and-suspenders guard for direct
-    // RPC bypass (CLI testing, future plugin, etc.) so the engine never loads
-    // weights against a tokenizer/config that has since been re-uploaded.
-    if state.outdated_models.lock().await.contains(&active_id) {
-        return Err(format!(
-            "active model `{active_id}` is out of date — open the MODELS card \
-             and click Update first, then start the server"
-        ));
+    // active chat model as out of date.  The frontend already greys out the
+    // Start button in that state — this is the belt-and-suspenders guard for
+    // direct RPC bypass (CLI testing, future plugin, etc.) so the engine never
+    // loads weights against a tokenizer/config that has since been re-uploaded.
+    if let Some(ref cid) = chat_id {
+        if state.outdated_models.lock().await.contains(cid) {
+            return Err(format!(
+                "active model `{cid}` is out of date — open the MODELS card \
+                 and click Update first, then start the server"
+            ));
+        }
     }
 
-    // Resolve the on-disk path of the active model. Flat-layout dirs use the
-    // dir name (e.g. `gemma-4-26b-a4b-mlx-imatrix3plus-awq`) which won't match the HF Hub
-    // id (`hsng95/gemma-4-26b-a4b-mlx-imatrix3plus-awq`). When the model is found
-    // locally we pass the absolute path as MODEL_ID — the MLX native runner
-    // and tokenizer loader both detect `is_dir()` and skip HF Hub entirely.
+    // Resolve the on-disk path of the active chat model. Flat-layout dirs use
+    // the dir name (e.g. `gemma-4-26b-a4b-mlx-imatrix3plus-awq`) which won't
+    // match the HF Hub id (`hsng95/gemma-4-26b-a4b-mlx-imatrix3plus-awq`). When
+    // the model is found locally we pass the absolute path as MODEL_ID — the MLX
+    // native runner and tokenizer loader both detect `is_dir()` and skip HF Hub
+    // entirely. The diffusion backend resolves its own component repos from
+    // IMAGE_MODEL_ID, so the image slot needs no path resolution here.
     let cat = state.catalog.lock().await;
-    // Diffusion image model → launch the server in dedicated image mode and skip
-    // the chat-model path/byte resolution (the diffusion backend loads its own
-    // hardcoded component repos; MODEL_ID is unused in image mode).
-    if cat.is_image_model(&active_id) {
-        drop(cat);
-        return sup
-            .start(app, &cfg, &active_id, None, /* image_mode */ true)
-            .await
-            .map_err(err);
-    }
-    let entries = models::scan_local(&models_dir, &cat).map_err(err)?;
-    let active_entry = entries.iter().find(|m| m.id == active_id);
-
-    let (model_arg, active_bytes) = if let Some(entry) = active_entry {
-        // Mirror into local_model_dir too — engine.rs reads LUMEN_GEMMA4_DIR /
-        // LUMEN_QWEN35_SHARDS for the non-MLX (Candle) Gemma4Native /
-        // Qwen35Moe paths.
-        if cfg.server.local_model_dir.is_none() {
-            cfg.server.local_model_dir = Some(entry.path.clone());
+    let (model_arg, active_bytes) = if let Some(ref cid) = chat_id {
+        let entries = models::scan_local(&models_dir, &cat).map_err(err)?;
+        match entries.iter().find(|m| m.id == *cid) {
+            Some(entry) => {
+                // Mirror into local_model_dir too — engine.rs reads
+                // LUMEN_GEMMA4_DIR / LUMEN_QWEN35_SHARDS for the non-MLX
+                // (Candle) Gemma4Native / Qwen35Moe paths.
+                if cfg.server.local_model_dir.is_none() {
+                    cfg.server.local_model_dir = Some(entry.path.clone());
+                }
+                (
+                    entry.path.to_string_lossy().into_owned(),
+                    Some(entry.size_bytes),
+                )
+            }
+            None => {
+                // Not on disk — resolve to the canonical HF id so the server
+                // can fetch.
+                let id = cat
+                    .find_recommended(cid)
+                    .map(|r| r.id.clone())
+                    .unwrap_or_else(|| cid.clone());
+                (id, None)
+            }
         }
-        (
-            entry.path.to_string_lossy().into_owned(),
-            Some(entry.size_bytes),
-        )
     } else {
-        // Not on disk — resolve to the canonical HF id so the server can fetch.
-        let id = cat
-            .find_recommended(&active_id)
-            .map(|r| r.id.clone())
-            .unwrap_or_else(|| active_id.clone());
-        (id, None)
+        // Image-only: no LLM to resolve. Empty MODEL_ID → server omits it.
+        (String::new(), None)
     };
     drop(cat);
 
+    let serve = match (chat_id.is_some(), image_id.is_some()) {
+        (true, true) => server::ServeKind::Hybrid,
+        (false, true) => server::ServeKind::Image,
+        (true, false) => server::ServeKind::Chat,
+        // Guarded by the early `is_none() && is_none()` return above.
+        (false, false) => unreachable!(),
+    };
+
     sup.start(
         app,
         &cfg,
         &model_arg,
         active_bytes,
-        /* image_mode */ false,
+        image_id.as_deref(),
+        serve,
     )
     .await
     .map_err(err)

crates/lumen-app/src/config.rs

@@ -35,7 +35,16 @@ pub struct PersistentConfig {
     /// write wins, with a warning displayed in the UI).
     #[serde(default)]
     pub env_overrides: BTreeMap<String, String>,
+    /// Active chat / LLM model id. Independent of `active_image_model` so a
+    /// chat model and an image model can be selected at once (→ hybrid serve).
     pub active_model: Option<String>,
+    /// Active text-to-image diffusion model id (e.g. `flux2-dev`). Kept in a
+    /// separate slot from `active_model` so selecting an image model no longer
+    /// overwrites the chat model — with both set, the server launches in
+    /// `LUMEN_SERVE=hybrid` (LLM + diffusion co-resident). `#[serde(default)]`
+    /// keeps older configs (which lack the field) loadable.
+    #[serde(default)]
+    pub active_image_model: Option<String>,
     /// Optional override for the `lumen-server` binary path. When `None` the
     /// app searches PATH and (in bundled builds) the sidecar location.
     pub server_binary_path: Option<PathBuf>,
@@ -296,6 +305,7 @@ impl Default for PersistentConfig {
             },
             env_overrides: BTreeMap::new(),
             active_model: None,
+            active_image_model: None,
             server_binary_path: None,
             models_dir,
         }