diff --git a/dashboard/src/lib/components/FamilyLogos.svelte b/dashboard/src/lib/components/FamilyLogos.svelte
index 306a447ed9..e741bedb22 100644
--- a/dashboard/src/lib/components/FamilyLogos.svelte
+++ b/dashboard/src/lib/components/FamilyLogos.svelte
@@ -7,7 +7,14 @@
   let { family, class: className = "" }: FamilyLogoProps = $props();
 </script>
 
-{#if family === "favorites"}
+{#if family === "recommended"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M11 3l1.8 4.7L17.5 9.5l-4.7 1.8L11 16l-1.8-4.7L4.5 9.5l4.7-1.8L11 3z"
+    />
+    <path d="M18 13l.9 2.4 2.4.9-2.4.9-.9 2.4-.9-2.4-2.4-.9 2.4-.9.9-2.4z" />
+  </svg>
+{:else if family === "favorites"}
   <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
     <path
       d="M12 2l3.09 6.26L22 9.27l-5 4.87 1.18 6.88L12 17.77l-6.18 3.25L7 14.14 2 9.27l6.91-1.01L12 2z"
diff --git a/dashboard/src/lib/components/FamilySidebar.svelte b/dashboard/src/lib/components/FamilySidebar.svelte
index b90f777c9a..e153f0f051 100644
--- a/dashboard/src/lib/components/FamilySidebar.svelte
+++ b/dashboard/src/lib/components/FamilySidebar.svelte
@@ -4,6 +4,7 @@
   type FamilySidebarProps = {
     families: string[];
     selectedFamily: string | null;
+    hasRecommended: boolean;
     hasFavorites: boolean;
     hasRecents: boolean;
     onSelect: (family: string | null) => void;
@@ -12,6 +13,7 @@
   let {
     families,
     selectedFamily,
+    hasRecommended,
     hasFavorites,
     hasRecents,
     onSelect,
@@ -19,6 +21,7 @@
 
   // Family display names
   const familyNames: Record<string, string> = {
+    recommended: "Recommended",
     favorites: "Favorites",
     recents: "Recent",
     huggingface: "Hub",
@@ -45,6 +48,31 @@
 <div
   class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[80px] sm:min-w-[72px] overflow-y-auto scrollbar-hide"
 >
+  <!-- Recommended (curated set) -->
+  {#if hasRecommended}
+    <button
+      type="button"
+      onclick={() => onSelect("recommended")}
+      class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {selectedFamily ===
+      'recommended'
+        ? 'bg-exo-yellow/20 border-l-2 border-exo-yellow'
+        : 'hover:bg-white/5 border-l-2 border-transparent'}"
+      title="Recommended models"
+    >
+      <FamilyLogos
+        family="recommended"
+        class={selectedFamily === "recommended"
+          ? "text-exo-yellow"
+          : "text-white/50 group-hover:text-exo-yellow/70"}
+      />
+      <span
+        class="text-[11px] font-mono mt-0.5 {selectedFamily === 'recommended'
+          ? 'text-exo-yellow'
+          : 'text-white/40 group-hover:text-white/60'}">Rec'd</span
+      >
+    </button>
+  {/if}
+
   <!-- All models (no filter) -->
   <button
     type="button"
diff --git a/dashboard/src/lib/components/ModelPickerModal.svelte b/dashboard/src/lib/components/ModelPickerModal.svelte
index b74fd1ba9e..8acbb4f8fc 100644
--- a/dashboard/src/lib/components/ModelPickerModal.svelte
+++ b/dashboard/src/lib/components/ModelPickerModal.svelte
@@ -22,6 +22,7 @@
     is_custom?: boolean;
     tasks?: string[];
     hugging_face_id?: string;
+    recommended?: boolean;
   }
 
   interface ModelGroup {
@@ -110,7 +111,7 @@
 
   // Local state
   let searchQuery = $state("");
-  let selectedFamily = $state<string | null>(null);
+  let selectedFamily = $state<string | null>("recommended");
   let expandedGroups = $state<Set<string>>(new Set());
   let showFilters = $state(false);
   let filters = $state<FilterState>({
@@ -210,6 +211,9 @@
       manualModelId = "";
       addModelError = null;
       justAddedModelId = null;
+      if (selectedFamily === "recommended" && !hasRecommended) {
+        selectedFamily = null;
+      }
       if (justAddedTimer) {
         clearTimeout(justAddedTimer);
         justAddedTimer = null;
@@ -242,6 +246,7 @@
       selectedFamily === "huggingface" ||
       selectedFamily === "recents" ||
       selectedFamily === "favorites" ||
+      selectedFamily === "recommended" ||
       query.length < 2 ||
       !noLocalResults
     ) {
@@ -477,7 +482,20 @@
     let result: ModelGroup[] = [...groupedModels];
 
     // Filter by family
-    if (selectedFamily === "favorites") {
+    if (selectedFamily === "recommended") {
+      result = result
+        .map((g) => {
+          const variants = g.variants.filter((v) => v.recommended);
+          if (variants.length === 0) return null;
+          return {
+            ...g,
+            variants,
+            smallestVariant: variants[0],
+            hasMultipleVariants: variants.length > 1,
+          };
+        })
+        .filter((g): g is ModelGroup => g !== null);
+    } else if (selectedFamily === "favorites") {
       result = result.filter((g) => favorites.has(g.id));
     } else if (
       selectedFamily &&
@@ -567,6 +585,10 @@
   // Check if any favorites exist
   const hasFavorites = $derived(favorites.size > 0);
 
+  const hasRecommended = $derived(
+    groupedModels.some((g) => g.variants.some((v) => v.recommended)),
+  );
+
   // Timestamp lookup for recent models
   const recentTimestamps = $derived(
     new Map(getRecentEntries().map((e) => [e.modelId, e.launchedAt])),
@@ -609,8 +631,9 @@
     );
   });
 
-  // Split filtered groups into recommended (fits_now) and others for visual separation
-  const recommendedGroups = $derived(
+  // Split filtered groups into ones that fit now and others, for visual
+  // separation (distinct from the curated "Recommended" tab).
+  const fitsGroups = $derived(
     filteredGroups.filter((g) =>
       g.variants.some((v) => getModelFitStatus(v.id) === "fits_now"),
     ),
@@ -783,6 +806,7 @@
       <FamilySidebar
         families={uniqueFamilies}
         {selectedFamily}
+        {hasRecommended}
         {hasFavorites}
         hasRecents={hasRecentsTab}
         onSelect={(family) => (selectedFamily = family)}
@@ -958,8 +982,8 @@
             {/if}
           </div>
         {:else}
-          <!-- Recommended for your cluster -->
-          {#if recommendedGroups.length > 0 && otherGroups.length > 0 && !searchQuery.trim()}
+          <!-- Fits in available memory -->
+          {#if fitsGroups.length > 0 && otherGroups.length > 0 && !searchQuery.trim()}
             <div
               class="sticky top-0 z-10 flex items-center gap-2 px-3 py-2 bg-green-950/60 border-b border-green-500/20 backdrop-blur-sm"
             >
@@ -978,14 +1002,11 @@
               </svg>
               <span
                 class="text-xs font-mono text-green-400 tracking-wider uppercase"
-                >Recommended for your cluster</span
-              >
-              <span class="text-xs font-mono text-green-400/50"
-                >— fits in available memory</span
+                >Fits in available memory</span
               >
             </div>
           {/if}
-          {#each recommendedGroups as group}
+          {#each fitsGroups as group}
             <ModelPickerGroup
               {group}
               isExpanded={expandedGroups.has(group.id)}
@@ -1004,7 +1025,7 @@
             />
           {/each}
           <!-- Other models -->
-          {#if otherGroups.length > 0 && recommendedGroups.length > 0 && !searchQuery.trim()}
+          {#if otherGroups.length > 0 && fitsGroups.length > 0 && !searchQuery.trim()}
             <div
               class="sticky top-0 z-10 flex items-center gap-2 px-3 py-2 bg-exo-dark-gray/80 border-y border-exo-medium-gray/20 backdrop-blur-sm"
             >
diff --git a/dashboard/src/lib/components/PrefillDecodeDisaggregation.svelte b/dashboard/src/lib/components/PrefillDecodeDisaggregation.svelte
index 7b9c1f6558..1e32005447 100644
--- a/dashboard/src/lib/components/PrefillDecodeDisaggregation.svelte
+++ b/dashboard/src/lib/components/PrefillDecodeDisaggregation.svelte
@@ -1,5 +1,5 @@
 <script lang="ts">
-  import { onMount, onDestroy } from "svelte";
+  import { onMount } from "svelte";
   import FamilyLogos from "$lib/components/FamilyLogos.svelte";
   import {
     instances,
@@ -19,14 +19,10 @@
     VllmInstance?: Instance;
   };
 
-  let interval: ReturnType<typeof setInterval> | null = null;
-
   onMount(() => {
+    // One immediate refresh on mount; live updates ride the global /state poll
+    // in app.svelte.ts (no separate interval here — it doubled /state traffic).
     refreshState();
-    interval = setInterval(refreshState, 3000);
-  });
-  onDestroy(() => {
-    if (interval) clearInterval(interval);
   });
 
   type InstanceRow = {
diff --git a/dashboard/src/lib/stores/app.svelte.ts b/dashboard/src/lib/stores/app.svelte.ts
index a717c3afbe..2e33755c7c 100644
--- a/dashboard/src/lib/stores/app.svelte.ts
+++ b/dashboard/src/lib/stores/app.svelte.ts
@@ -606,9 +606,12 @@ class AppStore {
   /** Number of consecutive fetch failures. */
   private consecutiveFailures = 0;
   private static readonly CONNECTION_LOST_THRESHOLD = 3;
+  private static readonly STATE_POLL_INTERVAL_MS = 2000;
 
   private fetchInterval: ReturnType<typeof setInterval> | null = null;
   private previewsInterval: ReturnType<typeof setInterval> | null = null;
+  private stateEtag: string | null = null;
+  private visibilityListener: (() => void) | null = null;
   private lastConversationPersistTs = 0;
   private previousNodeIds: Set<string> = new Set();
 
@@ -1284,14 +1287,45 @@ class AppStore {
   startPolling() {
     this.fetchState();
     this.fetchFeatureFlags();
-    this.fetchInterval = setInterval(() => this.fetchState(), 1000);
+    this.startStateInterval();
+
+    // Pause polling while the tab is hidden — a backgrounded dashboard has no
+    // reason to keep pulling /state. Resume with an immediate fetch when the
+    // tab becomes visible again.
+    if (typeof document !== "undefined" && !this.visibilityListener) {
+      this.visibilityListener = () => {
+        if (document.hidden) {
+          this.stopStateInterval();
+        } else if (!this.fetchInterval) {
+          this.fetchState();
+          this.startStateInterval();
+        }
+      };
+      document.addEventListener("visibilitychange", this.visibilityListener);
+    }
   }
 
-  stopPolling() {
+  private startStateInterval() {
+    if (this.fetchInterval) return;
+    this.fetchInterval = setInterval(
+      () => this.fetchState(),
+      AppStore.STATE_POLL_INTERVAL_MS,
+    );
+  }
+
+  private stopStateInterval() {
     if (this.fetchInterval) {
       clearInterval(this.fetchInterval);
       this.fetchInterval = null;
     }
+  }
+
+  stopPolling() {
+    this.stopStateInterval();
+    if (this.visibilityListener && typeof document !== "undefined") {
+      document.removeEventListener("visibilitychange", this.visibilityListener);
+      this.visibilityListener = null;
+    }
     this.stopPreviewsPolling();
   }
 
@@ -1307,10 +1341,25 @@ class AppStore {
 
   async fetchState() {
     try {
-      const response = await fetch("/state");
+      const headers: Record<string, string> = {};
+      if (this.stateEtag) {
+        headers["If-None-Match"] = this.stateEtag;
+      }
+      const response = await fetch("/state", { headers });
+
+      // 304: state unchanged since our last poll — no body shipped. Keep the
+      // data we already have and treat it as a successful poll.
+      if (response.status === 304) {
+        this.lastUpdate = Date.now();
+        if (!this.isConnected) this.isConnected = true;
+        this.consecutiveFailures = 0;
+        return;
+      }
+
       if (!response.ok) {
         throw new Error(`Failed to fetch state: ${response.status}`);
       }
+      this.stateEtag = response.headers.get("ETag");
       const data: RawStateResponse = await response.json();
 
       if (data.topology) {
diff --git a/dashboard/src/routes/+page.svelte b/dashboard/src/routes/+page.svelte
index 5db620795e..bd3564b011 100644
--- a/dashboard/src/routes/+page.svelte
+++ b/dashboard/src/routes/+page.svelte
@@ -783,6 +783,7 @@
       quantization?: string;
       base_model?: string;
       capabilities?: string[];
+      recommended?: boolean;
     }>
   >([]);
   type ModelMemoryFitStatus =
diff --git a/resources/inference_model_cards/mlx-community--DeepSeek-V4-Flash.toml b/resources/inference_model_cards/mlx-community--DeepSeek-V4-Flash.toml
index c4f64474a5..33b3012280 100644
--- a/resources/inference_model_cards/mlx-community--DeepSeek-V4-Flash.toml
+++ b/resources/inference_model_cards/mlx-community--DeepSeek-V4-Flash.toml
@@ -12,6 +12,7 @@ reasoning_dialect = "tool_conditional"
 
 context_length = 1048576
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 155095760030
 
diff --git a/resources/inference_model_cards/mlx-community--DeepSeek-V4-Pro.toml b/resources/inference_model_cards/mlx-community--DeepSeek-V4-Pro.toml
index 00dd3e66e0..ac18fc3cc4 100644
--- a/resources/inference_model_cards/mlx-community--DeepSeek-V4-Pro.toml
+++ b/resources/inference_model_cards/mlx-community--DeepSeek-V4-Pro.toml
@@ -12,6 +12,7 @@ reasoning_dialect = "tool_conditional"
 
 context_length = 1048576
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 849681803879
 
diff --git a/resources/inference_model_cards/mlx-community--GLM-5.1-DQ4plus-q8.toml b/resources/inference_model_cards/mlx-community--GLM-5.1-DQ4plus-q8.toml
index c92ffd2524..85124a61b6 100644
--- a/resources/inference_model_cards/mlx-community--GLM-5.1-DQ4plus-q8.toml
+++ b/resources/inference_model_cards/mlx-community--GLM-5.1-DQ4plus-q8.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "post_last_user"
 context_length = 202752
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 465173655552
 
diff --git a/resources/inference_model_cards/mlx-community--GLM-5.1-MXFP4-Q8.toml b/resources/inference_model_cards/mlx-community--GLM-5.1-MXFP4-Q8.toml
index 8230d2fdb4..c717182e15 100644
--- a/resources/inference_model_cards/mlx-community--GLM-5.1-MXFP4-Q8.toml
+++ b/resources/inference_model_cards/mlx-community--GLM-5.1-MXFP4-Q8.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "post_last_user"
 context_length = 202752
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 405480321024
 
diff --git a/resources/inference_model_cards/mlx-community--GLM-5.1.toml b/resources/inference_model_cards/mlx-community--GLM-5.1.toml
index a59dddfb7a..22978d76a6 100644
--- a/resources/inference_model_cards/mlx-community--GLM-5.1.toml
+++ b/resources/inference_model_cards/mlx-community--GLM-5.1.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "post_last_user"
 context_length = 202752
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 1487822475264
 
diff --git a/resources/inference_model_cards/mlx-community--Kimi-K2.6-mlx-DQ3_K_M-q8.toml b/resources/inference_model_cards/mlx-community--Kimi-K2.6-mlx-DQ3_K_M-q8.toml
index b2444d7d2a..38cfbe642c 100644
--- a/resources/inference_model_cards/mlx-community--Kimi-K2.6-mlx-DQ3_K_M-q8.toml
+++ b/resources/inference_model_cards/mlx-community--Kimi-K2.6-mlx-DQ3_K_M-q8.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "suffix"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 470628683776
 
diff --git a/resources/inference_model_cards/mlx-community--Llama-3.2-1B-Instruct-4bit.toml b/resources/inference_model_cards/mlx-community--Llama-3.2-1B-Instruct-4bit.toml
index 4a70c63a35..69090c4c86 100644
--- a/resources/inference_model_cards/mlx-community--Llama-3.2-1B-Instruct-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--Llama-3.2-1B-Instruct-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text"]
 
 context_length = 131072
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 729808896
 
diff --git a/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit-mxfp4.toml b/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit-mxfp4.toml
index 2759c97f12..da03b13edb 100644
--- a/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit-mxfp4.toml
+++ b/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit-mxfp4.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "post_last_user"
 context_length = 196608
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 121537496794
 
diff --git a/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit.toml b/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit.toml
index 8d1a5ad250..761b8c35ad 100644
--- a/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--MiniMax-M2.7-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "post_last_user"
 context_length = 196608
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 128682598717
 
diff --git a/resources/inference_model_cards/mlx-community--MiniMax-M2.7-8bit.toml b/resources/inference_model_cards/mlx-community--MiniMax-M2.7-8bit.toml
index 97ce9a75b8..1847dc73c6 100644
--- a/resources/inference_model_cards/mlx-community--MiniMax-M2.7-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--MiniMax-M2.7-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "post_last_user"
 context_length = 196608
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 243002680786
 
diff --git a/resources/inference_model_cards/mlx-community--MiniMax-M2.7.toml b/resources/inference_model_cards/mlx-community--MiniMax-M2.7.toml
index e5acfbdab5..a4d146175a 100644
--- a/resources/inference_model_cards/mlx-community--MiniMax-M2.7.toml
+++ b/resources/inference_model_cards/mlx-community--MiniMax-M2.7.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "post_last_user"
 context_length = 196608
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 457492783366
 
diff --git a/resources/inference_model_cards/mlx-community--Nemotron-3-Ultra-550B-A55B-4bit.toml b/resources/inference_model_cards/mlx-community--Nemotron-3-Ultra-550B-A55B-4bit.toml
new file mode 100644
index 0000000000..0b3b48ba05
--- /dev/null
+++ b/resources/inference_model_cards/mlx-community--Nemotron-3-Ultra-550B-A55B-4bit.toml
@@ -0,0 +1,21 @@
+model_id = "mlx-community/Nemotron-3-Ultra-550B-A55B-4bit"
+n_layers = 108
+hidden_size = 8192
+num_key_value_heads = 2
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "4bit"
+base_model = "NVIDIA Nemotron-3-Ultra-550B-A55B"
+capabilities = ["text"]
+
+context_length = 262144
+backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
+[storage_size]
+in_bytes = 346976565922
+
+# Source: https://huggingface.co/mlx-community/Nemotron-3-Ultra-550B-A55B-4bit
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
diff --git a/resources/inference_model_cards/mlx-community--Nemotron-3-Ultra-550B-A55B.toml b/resources/inference_model_cards/mlx-community--Nemotron-3-Ultra-550B-A55B.toml
new file mode 100644
index 0000000000..31d62eb65c
--- /dev/null
+++ b/resources/inference_model_cards/mlx-community--Nemotron-3-Ultra-550B-A55B.toml
@@ -0,0 +1,21 @@
+model_id = "mlx-community/Nemotron-3-Ultra-550B-A55B"
+n_layers = 108
+hidden_size = 8192
+num_key_value_heads = 2
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "nemotron"
+quantization = "int4"
+base_model = "NVIDIA Nemotron-3-Ultra-550B-A55B"
+capabilities = ["text"]
+
+context_length = 262144
+backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
+[storage_size]
+in_bytes = 360783261541
+
+# Source: https://huggingface.co/mlx-community/Nemotron-3-Ultra-550B-A55B
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
diff --git a/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-4bit.toml b/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-4bit.toml
index d692f7dc62..1b9908d05c 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "code"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 45644286500
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-8bit.toml b/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-8bit.toml
index 8ecf357fec..e9d2092975 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "code"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 89357758772
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-bf16.toml b/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-bf16.toml
index 73c32b0ef7..1231948be9 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-bf16.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3-Coder-Next-bf16.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "code"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 157548627945
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-4bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-4bit.toml
index 5e50aca6d2..902267c03b 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 69593314272
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-8bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-8bit.toml
index ede2e81c2b..3828fd3314 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 130648036320
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-bf16.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-bf16.toml
index e1cdb9f5f3..5dac5984b1 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-bf16.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-122B-A10B-bf16.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 245125640160
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-2B-MLX-8bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-2B-MLX-8bit.toml
index 78ff87479e..873b8812ba 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-2B-MLX-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-2B-MLX-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 2662787264
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-4bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-4bit.toml
index 202b828616..28dcb694f3 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 223860768352
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-8bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-8bit.toml
index a0cef3ff97..f7298452e4 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-397B-A17B-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 422032580192
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-9B-4bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-9B-4bit.toml
index 2248f19364..f861f1b631 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-9B-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-9B-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 5950062560
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.5-9B-8bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.5-9B-8bit.toml
index 44cbc45611..4e3f7ae093 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.5-9B-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.5-9B-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 10426433504
 
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.6-27B-4bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.6-27B-4bit.toml
index 88decd0033..edf13d5298 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.6-27B-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.6-27B-4bit.toml
@@ -11,18 +11,19 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 16054262240
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-27B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
 [sampling_defaults]
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
 top_k = 20
 min_p = 0.0
 repetition_penalty = 1.0
-presence_penalty = 1.5
+presence_penalty = 0.0
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-27B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.6-27B-8bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.6-27B-8bit.toml
index 8cef7490a2..7899e3fe91 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.6-27B-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.6-27B-8bit.toml
@@ -11,18 +11,19 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 29500938720
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-27B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
 [sampling_defaults]
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
 top_k = 20
 min_p = 0.0
 repetition_penalty = 1.0
-presence_penalty = 1.5
+presence_penalty = 0.0
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-27B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.6-27B-bf16.toml b/resources/inference_model_cards/mlx-community--Qwen3.6-27B-bf16.toml
index d04042daac..63e4cff12c 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.6-27B-bf16.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.6-27B-bf16.toml
@@ -11,18 +11,19 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 54713457120
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-27B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
 [sampling_defaults]
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
 top_k = 20
 min_p = 0.0
 repetition_penalty = 1.0
-presence_penalty = 1.5
+presence_penalty = 0.0
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-27B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-4bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-4bit.toml
index d2ca90e8f9..a47bf3fdb3 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-4bit.toml
@@ -11,18 +11,19 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 20401929952
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
 [sampling_defaults]
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
 top_k = 20
 min_p = 0.0
 repetition_penalty = 1.0
-presence_penalty = 1.5
+presence_penalty = 0.0
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-5bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-5bit.toml
index 8229424f65..55213afcf5 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-5bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-5bit.toml
@@ -17,12 +17,12 @@ in_bytes = 24731729632
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
 [sampling_defaults]
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
 top_k = 20
 min_p = 0.0
 repetition_penalty = 1.0
-presence_penalty = 1.5
+presence_penalty = 0.0
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-8bit.toml b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-8bit.toml
index c03748fae3..5839066c79 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-8bit.toml
@@ -11,18 +11,19 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 37721128672
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
 [sampling_defaults]
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
 top_k = 20
 min_p = 0.0
 repetition_penalty = 1.0
-presence_penalty = 1.5
+presence_penalty = 0.0
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
diff --git a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-bf16.toml b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-bf16.toml
index 9bd5cf1ebe..0a174730ac 100644
--- a/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-bf16.toml
+++ b/resources/inference_model_cards/mlx-community--Qwen3.6-35B-A3B-bf16.toml
@@ -11,18 +11,19 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 reasoning_dialect = "post_last_user"
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 70214363872
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
 [sampling_defaults]
-temperature = 1.0
+temperature = 0.6
 top_p = 0.95
 top_k = 20
 min_p = 0.0
 repetition_penalty = 1.0
-presence_penalty = 1.5
+presence_penalty = 0.0
 
 # Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B#best-practices
 # Source: https://unsloth.ai/docs/models/qwen3.5
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-12B-it-4bit.toml b/resources/inference_model_cards/mlx-community--gemma-4-12B-it-4bit.toml
new file mode 100644
index 0000000000..da341d04ea
--- /dev/null
+++ b/resources/inference_model_cards/mlx-community--gemma-4-12B-it-4bit.toml
@@ -0,0 +1,23 @@
+model_id = "mlx-community/gemma-4-12B-it-4bit"
+n_layers = 48
+hidden_size = 3840
+num_key_value_heads = 8
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "gemma"
+quantization = "4bit"
+base_model = "Gemma 4 12B"
+capabilities = ["text", "vision"]
+
+context_length = 131072
+backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
+[storage_size]
+in_bytes = 11020140510
+
+# Source: https://ai.google.dev/gemma/docs/core/model_card_4
+# Source: https://huggingface.co/mlx-community/gemma-4-12B-it-4bit
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
+top_k = 64
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-12B-it-8bit.toml b/resources/inference_model_cards/mlx-community--gemma-4-12B-it-8bit.toml
new file mode 100644
index 0000000000..6e85d70baa
--- /dev/null
+++ b/resources/inference_model_cards/mlx-community--gemma-4-12B-it-8bit.toml
@@ -0,0 +1,23 @@
+model_id = "mlx-community/gemma-4-12B-it-8bit"
+n_layers = 48
+hidden_size = 3840
+num_key_value_heads = 8
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "gemma"
+quantization = "8bit"
+base_model = "Gemma 4 12B"
+capabilities = ["text", "vision"]
+
+context_length = 131072
+backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
+[storage_size]
+in_bytes = 12748570647
+
+# Source: https://ai.google.dev/gemma/docs/core/model_card_4
+# Source: https://huggingface.co/mlx-community/gemma-4-12B-it-8bit
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
+top_k = 64
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-12B-it-bf16.toml b/resources/inference_model_cards/mlx-community--gemma-4-12B-it-bf16.toml
new file mode 100644
index 0000000000..f62afc5000
--- /dev/null
+++ b/resources/inference_model_cards/mlx-community--gemma-4-12B-it-bf16.toml
@@ -0,0 +1,23 @@
+model_id = "mlx-community/gemma-4-12B-it-bf16"
+n_layers = 48
+hidden_size = 3840
+num_key_value_heads = 8
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "gemma"
+quantization = "bf16"
+base_model = "Gemma 4 12B"
+capabilities = ["text", "vision"]
+
+context_length = 131072
+backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
+[storage_size]
+in_bytes = 23951814902
+
+# Source: https://ai.google.dev/gemma/docs/core/model_card_4
+# Source: https://huggingface.co/mlx-community/gemma-4-12B-it-bf16
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
+top_k = 64
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-4bit.toml b/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-4bit.toml
index 729b4c5eed..ea100f43af 100644
--- a/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "vision"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 15608614044
 
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-8bit.toml b/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-8bit.toml
index 36ac44ddff..f260e48207 100644
--- a/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "vision"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 27953417372
 
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-bf16.toml b/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-bf16.toml
index e119725444..2f5caab014 100644
--- a/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-bf16.toml
+++ b/resources/inference_model_cards/mlx-community--gemma-4-26b-a4b-it-bf16.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "vision"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 51611872412
 
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-31b-it-4bit.toml b/resources/inference_model_cards/mlx-community--gemma-4-31b-it-4bit.toml
index 409db19731..d31b7ef77a 100644
--- a/resources/inference_model_cards/mlx-community--gemma-4-31b-it-4bit.toml
+++ b/resources/inference_model_cards/mlx-community--gemma-4-31b-it-4bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "vision"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 18411755224
 
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-31b-it-8bit.toml b/resources/inference_model_cards/mlx-community--gemma-4-31b-it-8bit.toml
index 378447a469..b3d615094e 100644
--- a/resources/inference_model_cards/mlx-community--gemma-4-31b-it-8bit.toml
+++ b/resources/inference_model_cards/mlx-community--gemma-4-31b-it-8bit.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "vision"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 33762858712
 
diff --git a/resources/inference_model_cards/mlx-community--gemma-4-31b-it-bf16.toml b/resources/inference_model_cards/mlx-community--gemma-4-31b-it-bf16.toml
index ea3c87d8be..f5862ab01e 100644
--- a/resources/inference_model_cards/mlx-community--gemma-4-31b-it-bf16.toml
+++ b/resources/inference_model_cards/mlx-community--gemma-4-31b-it-bf16.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "vision"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 62546177752
 
diff --git a/resources/inference_model_cards/mlx-community--gpt-oss-120b-MXFP4-Q8.toml b/resources/inference_model_cards/mlx-community--gpt-oss-120b-MXFP4-Q8.toml
index 091b079ba6..02c5936edc 100644
--- a/resources/inference_model_cards/mlx-community--gpt-oss-120b-MXFP4-Q8.toml
+++ b/resources/inference_model_cards/mlx-community--gpt-oss-120b-MXFP4-Q8.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "channel"
 context_length = 131072
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 70652212224
 
diff --git a/resources/inference_model_cards/mlx-community--gpt-oss-20b-MXFP4-Q8.toml b/resources/inference_model_cards/mlx-community--gpt-oss-20b-MXFP4-Q8.toml
index 9e1a56bae7..75228f8f24 100644
--- a/resources/inference_model_cards/mlx-community--gpt-oss-20b-MXFP4-Q8.toml
+++ b/resources/inference_model_cards/mlx-community--gpt-oss-20b-MXFP4-Q8.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking"]
 reasoning_dialect = "channel"
 context_length = 131072
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 12025908224
 
diff --git a/resources/inference_model_cards/moonshotai--Kimi-K2.6.toml b/resources/inference_model_cards/moonshotai--Kimi-K2.6.toml
index b169e51ae4..9f4f4d3914 100644
--- a/resources/inference_model_cards/moonshotai--Kimi-K2.6.toml
+++ b/resources/inference_model_cards/moonshotai--Kimi-K2.6.toml
@@ -11,6 +11,7 @@ capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 
 context_length = 262144
 backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+recommended = true
 [storage_size]
 in_bytes = 595148192736
 
diff --git a/src/exo/api/main.py b/src/exo/api/main.py
index fcd54c9315..7891b05792 100644
--- a/src/exo/api/main.py
+++ b/src/exo/api/main.py
@@ -1,5 +1,6 @@
 import base64
 import contextlib
+import gzip
 import hashlib
 import json
 import random
@@ -13,7 +14,16 @@
 
 import anyio
 from anyio import BrokenResourceError, ClosedResourceError
-from fastapi import FastAPI, File, Form, HTTPException, Query, Request, UploadFile
+from fastapi import (
+    FastAPI,
+    File,
+    Form,
+    HTTPException,
+    Query,
+    Request,
+    Response,
+    UploadFile,
+)
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, JSONResponse, StreamingResponse
 from fastapi.staticfiles import StaticFiles
@@ -22,6 +32,8 @@
 from hypercorn.typing import ASGIFramework
 from hypercorn.utils import LifespanTimeoutError, ShutdownError
 from loguru import logger
+from starlette.datastructures import Headers, MutableHeaders
+from starlette.types import ASGIApp, Message, Receive, Scope, Send
 
 from exo.api.adapters.chat_completions import (
     chat_request_to_text_generation,
@@ -235,6 +247,74 @@ def _require_disaggregation_enabled() -> None:
         )
 
 
+class ContentTypeGZipMiddleware:
+    _COMPRESSIBLE: tuple[str, ...] = ("application/json", "text/plain")
+    _MIN_SIZE: int = 500
+
+    def __init__(self, app: ASGIApp) -> None:
+        self.app = app
+
+    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
+        if scope["type"] != "http":
+            await self.app(scope, receive, send)
+            return
+
+        headers = Headers(scope=scope)
+        if "gzip" not in headers.get("accept-encoding", ""):
+            await self.app(scope, receive, send)
+            return
+
+        start_message: Message | None = None
+        body_chunks: list[bytes] = []
+        compressible = False
+
+        async def buffering_send(message: Message) -> None:
+            nonlocal start_message, compressible
+            message_type: str = message["type"]  # pyright: ignore[reportAny]
+            if message_type == "http.response.start":
+                start_message = message
+                response_headers = MutableHeaders(raw=start_message["headers"])  # pyright: ignore[reportAny]
+                content_type = response_headers.get("content-type", "")
+                already_encoded = "content-encoding" in response_headers
+                compressible = (not already_encoded) and any(
+                    content_type.startswith(prefix) for prefix in self._COMPRESSIBLE
+                )
+                if not compressible:
+                    await send(message)
+                return
+
+            if message_type != "http.response.body":
+                await send(message)
+                return
+
+            if not compressible:
+                await send(message)
+                return
+
+            assert start_message is not None
+            body_chunks.append(message.get("body", b""))  # pyright: ignore[reportAny]
+            if message.get("more_body", False):
+                return
+
+            body = b"".join(body_chunks)
+            response_headers = MutableHeaders(raw=start_message["headers"])  # pyright: ignore[reportAny]
+            if len(body) >= self._MIN_SIZE:
+                body = gzip.compress(body)
+                response_headers["content-encoding"] = "gzip"
+                response_headers["content-length"] = str(len(body))
+                vary = response_headers.get("vary")
+                if not vary:
+                    response_headers["vary"] = "Accept-Encoding"
+                elif "accept-encoding" not in vary.lower():
+                    response_headers["vary"] = f"{vary}, Accept-Encoding"
+            else:
+                response_headers["content-length"] = str(len(body))
+            await send(start_message)
+            await send({"type": "http.response.body", "body": body, "more_body": False})
+
+        await self.app(scope, receive, buffering_send)
+
+
 class API:
     def __init__(
         self,
@@ -274,6 +354,7 @@ async def _log_requests(  # pyright: ignore[reportUnusedFunction]
 
         self._setup_exception_handlers()
         self._setup_cors()
+        self._setup_gzip()
         self._setup_routes()
 
         self.app.mount(
@@ -340,6 +421,9 @@ def _setup_cors(self) -> None:
             allow_headers=["*"],
         )
 
+    def _setup_gzip(self) -> None:
+        self.app.add_middleware(ContentTypeGZipMiddleware)
+
     def _setup_routes(self) -> None:
         self.app.get("/node_id")(lambda: self.node_id)
         self.app.post("/instance")(self.create_instance)
@@ -408,8 +492,12 @@ def _setup_routes(self) -> None:
         self.app.get("/onboarding")(self.get_onboarding)
         self.app.post("/onboarding")(self.complete_onboarding)
 
-    def get_state(self, path: str = ""):
+    def get_state(self, request: Request, response: Response, path: str = ""):
         if path == "":
+            etag = f'"{self.state.last_event_applied_idx}"'
+            if request.headers.get("if-none-match") == etag:
+                return Response(status_code=304, headers={"ETag": etag})
+            response.headers["ETag"] = etag
             return self.state
         try:
             x = self.state.model_dump(by_alias=True)
@@ -1813,6 +1901,7 @@ async def get_models(self, status: str | None = Query(default=None)) -> ModelLis
                     capabilities=card.capabilities,
                     reasoning_dialect=card.reasoning_dialect,
                     context_length=card.context_length,
+                    recommended=card.recommended,
                 )
                 for card in cards
             ]
diff --git a/src/exo/api/types/api.py b/src/exo/api/types/api.py
index 02e397adaa..1929510f01 100644
--- a/src/exo/api/types/api.py
+++ b/src/exo/api/types/api.py
@@ -49,6 +49,7 @@ class ModelListModel(BaseModel):
     base_model: str = Field(default="")
     capabilities: list[str] = Field(default_factory=list)
     reasoning_dialect: ReasoningDialect = "none"
+    recommended: bool = Field(default=False)
 
 
 class ModelList(BaseModel):
diff --git a/src/exo/shared/apply.py b/src/exo/shared/apply.py
index 5f72247bac..5a630fb9f1 100644
--- a/src/exo/shared/apply.py
+++ b/src/exo/shared/apply.py
@@ -61,7 +61,9 @@
     NodeConfig,
     NodeDiskUsage,
     NodeNetworkInterfaces,
+    NvmlMetrics,
     RdmaCtlStatus,
+    SpbmMetrics,
     StaticNodeInformation,
     ThunderboltBridgeInfo,
 )
@@ -368,6 +370,29 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
                 event.node_id: info.system_profile,
             }
             update["node_memory"] = {**state.node_memory, event.node_id: info.memory}
+        case NvmlMetrics():
+            existing = state.node_system.get(event.node_id, info.system_profile)
+            nvml_update: dict[str, float] = {
+                "gpu_usage": info.system_profile.gpu_usage,
+                "temp": info.system_profile.temp,
+            }
+            if info.provides_sys_power:
+                nvml_update["sys_power"] = info.system_profile.sys_power
+            update["node_system"] = {
+                **state.node_system,
+                event.node_id: existing.model_copy(update=nvml_update),
+            }
+        case SpbmMetrics():
+            existing = state.node_system.get(event.node_id, info.system_profile)
+            update["node_system"] = {
+                **state.node_system,
+                event.node_id: existing.model_copy(
+                    update={
+                        "sys_power": info.system_profile.sys_power,
+                        "temp": info.system_profile.temp,
+                    }
+                ),
+            }
         case MemoryUsage():
             update["node_memory"] = {**state.node_memory, event.node_id: info}
         case NodeDiskUsage():
diff --git a/src/exo/shared/models/model_cards.py b/src/exo/shared/models/model_cards.py
index 8911b9323c..03eef683da 100644
--- a/src/exo/shared/models/model_cards.py
+++ b/src/exo/shared/models/model_cards.py
@@ -173,6 +173,7 @@ class ModelCard(FrozenModel):
     uses_cfg: bool = False
     trust_remote_code: bool = True
     is_custom: bool = False
+    recommended: bool = False
     vision: VisionCardConfig | None = None
     sampling_defaults: SamplingDefaults = Field(default_factory=SamplingDefaults)
 
diff --git a/src/exo/shared/types/profiling.py b/src/exo/shared/types/profiling.py
index a3548dc91a..44722290c3 100644
--- a/src/exo/shared/types/profiling.py
+++ b/src/exo/shared/types/profiling.py
@@ -39,6 +39,20 @@ def from_psutil(cls, *, override_memory: int | None) -> Self:
             swap_available=sm.free,
         )
 
+    @classmethod
+    def from_gpu_memory(
+        cls, *, gpu_total: int, gpu_free: int, override_memory: int | None
+    ) -> Self:
+        """Report GPU VRAM as the node's RAM (discrete GPUs like B300, where
+        host RAM misleads). Swap stays host-level — GPUs have none."""
+        sm = psutil.swap_memory()
+        return cls.from_bytes(
+            ram_total=gpu_total,
+            ram_available=gpu_free if override_memory is None else override_memory,
+            swap_total=sm.total,
+            swap_available=sm.free,
+        )
+
 
 class DiskUsage(FrozenModel):
     """Disk space usage for the models directory."""
diff --git a/src/exo/utils/info_gatherer/info_gatherer.py b/src/exo/utils/info_gatherer/info_gatherer.py
index 988d8b41f6..f8ba01b11b 100644
--- a/src/exo/utils/info_gatherer/info_gatherer.py
+++ b/src/exo/utils/info_gatherer/info_gatherer.py
@@ -32,6 +32,8 @@
 from exo.utils.task_group import TaskGroup
 
 from .macmon import MacmonMetrics
+from .nvml import NvmlMetrics, gather_gpu_memory, gather_nvidia_metrics, has_nvml
+from .spbm import SpbmMetrics, gather_spbm_metrics, has_spbm
 from .system_info import (
     get_friendly_name,
     get_model_and_chip,
@@ -356,7 +358,7 @@ async def _gather_iface_map() -> dict[str, str] | None:
 
 def _has_nvml_cuda() -> bool:
     try:
-        import pynvml as nvml  # pyright: ignore[reportMissingModuleSource]
+        import pynvml as nvml
     except ImportError:
         return False
     try:
@@ -385,6 +387,8 @@ async def gather(cls) -> Self:
 
 GatheredInfo = (
     MacmonMetrics
+    | NvmlMetrics
+    | SpbmMetrics
     | MemoryUsage
     | NodeNetworkInterfaces
     | MacThunderboltIdentifiers
@@ -450,6 +454,12 @@ async def run(self):
                 tg.start_soon(self._monitor_rdma_ctl_status, 10)
             if not IS_DARWIN:
                 tg.start_soon(self._monitor_memory_usage, 1)
+
+                spbm_available = await has_spbm()
+                if spbm_available:
+                    tg.start_soon(self._monitor_spbm_metrics, 0.5)
+                if has_nvml():
+                    tg.start_soon(self._monitor_nvml_metrics, 1, not spbm_available)
             tg.start_soon(self._watch_system_info, 10)
             tg.start_soon(self._monitor_misc, 60)
             tg.start_soon(self._monitor_static_info, 60)
@@ -520,9 +530,17 @@ async def _monitor_memory_usage(self, memory_poll_rate: float):
         )
         while True:
             try:
-                await self.info_sender.send(
-                    MemoryUsage.from_psutil(override_memory=override_memory)
-                )
+                gpu_memory = gather_gpu_memory()
+                if gpu_memory is not None:
+                    gpu_total, gpu_free = gpu_memory
+                    usage = MemoryUsage.from_gpu_memory(
+                        gpu_total=gpu_total,
+                        gpu_free=gpu_free,
+                        override_memory=override_memory,
+                    )
+                else:
+                    usage = MemoryUsage.from_psutil(override_memory=override_memory)
+                await self.info_sender.send(usage)
             except Exception as e:
                 logger.opt(exception=e).warning("Error gathering memory usage")
             await anyio.sleep(memory_poll_rate)
@@ -571,6 +589,28 @@ async def _monitor_disk_usage(self, disk_poll_interval: float):
                 logger.opt(exception=e).warning("Error gathering disk usage")
             await anyio.sleep(disk_poll_interval)
 
+    async def _monitor_nvml_metrics(
+        self, nvml_poll_rate: float, provides_sys_power: bool
+    ):
+        while True:
+            try:
+                metrics = gather_nvidia_metrics(provides_sys_power=provides_sys_power)
+                if metrics is not None:
+                    await self.info_sender.send(metrics)
+            except Exception as e:
+                logger.opt(exception=e).warning("Error gathering NVML metrics")
+            await anyio.sleep(nvml_poll_rate)
+
+    async def _monitor_spbm_metrics(self, spbm_poll_rate: float):
+        while True:
+            try:
+                metrics = await gather_spbm_metrics()
+                if metrics is not None:
+                    await self.info_sender.send(metrics)
+            except Exception as e:
+                logger.opt(exception=e).warning("Error gathering spbm metrics")
+            await anyio.sleep(spbm_poll_rate)
+
     async def _monitor_macmon(self, macmon_interval: float):
         if (
             macmon_path := os.getenv("EXO_MACMON_PATH") or shutil.which("macmon")
diff --git a/src/exo/utils/info_gatherer/nvml.py b/src/exo/utils/info_gatherer/nvml.py
new file mode 100644
index 0000000000..f25c932613
--- /dev/null
+++ b/src/exo/utils/info_gatherer/nvml.py
@@ -0,0 +1,159 @@
+import contextlib
+from typing import Protocol, cast
+
+from exo.shared.types.profiling import SystemPerformanceProfile
+from exo.utils.pydantic_ext import TaggedModel
+
+try:
+    import pynvml as nvml
+except ImportError:
+    nvml = None
+
+
+class _NvmlMemoryInfo(Protocol):
+    total: int
+    free: int
+
+
+_CPU_POWER_IDLE = 20.0
+_CPU_POWER_MAX = 100.0
+
+_DEVICE_GPU_MAX_WATTS: list[tuple[str, float]] = [
+    ("GB10", 100.0),
+]
+
+
+def _device_gpu_max_watts(name: str) -> float:
+    for substring, gpu_max_watts in _DEVICE_GPU_MAX_WATTS:
+        if substring in name:
+            return gpu_max_watts
+    return 0.0
+
+
+_UNIFIED_MEMORY_DEVICES: tuple[str, ...] = ("GB10",)
+
+
+class NvmlMetrics(TaggedModel):
+    system_profile: SystemPerformanceProfile
+    provides_sys_power: bool
+
+
+def has_nvml() -> bool:
+    if nvml is None:
+        return False
+    try:
+        nvml.nvmlInit()
+        count = nvml.nvmlDeviceGetCount()
+        nvml.nvmlShutdown()
+        return count > 0
+    except Exception:
+        return False
+
+
+def gather_gpu_memory() -> tuple[int, int] | None:
+    """Total and free GPU VRAM in bytes, summed across all NVIDIA devices.
+
+    Returns None when NVML is unavailable, no device is present, or the device
+    reports no usable VRAM — notably unified-memory devices (GB10/DGX Spark),
+    where NVML/nvidia-smi report no discrete memory pool (`memory.total` shows
+    [N/A]). Callers fall back to psutil host RAM in that case, which on a
+    unified device is the same pool and is the correct number anyway.
+
+    On discrete GPUs (e.g. B300) this is the meaningful capacity for model
+    load — host RAM under-reports/misleads there.
+    """
+    if nvml is None:
+        return None
+    is_init = False
+    try:
+        nvml.nvmlInit()
+        is_init = True
+        count = nvml.nvmlDeviceGetCount()
+        if count == 0:
+            return None
+        total = 0
+        free = 0
+        for i in range(count):
+            handle = nvml.nvmlDeviceGetHandleByIndex(i)
+            name = cast(str | bytes, nvml.nvmlDeviceGetName(handle))  # pyright: ignore[reportUnknownMemberType]
+            if isinstance(name, bytes):
+                name = name.decode("utf-8", errors="replace")
+            if any(substring in name for substring in _UNIFIED_MEMORY_DEVICES):
+                # Unified-memory device — host RAM (psutil) is the right number.
+                return None
+            mem = cast(_NvmlMemoryInfo, nvml.nvmlDeviceGetMemoryInfo(handle))  # pyright: ignore[reportUnknownMemberType]
+            total += int(mem.total)
+            free += int(mem.free)
+        if total <= 0:
+            return None
+        return total, free
+    except Exception:
+        return None
+    finally:
+        if is_init:
+            nvml.nvmlShutdown()
+
+
+def gather_nvidia_metrics(*, provides_sys_power: bool) -> NvmlMetrics | None:
+    if nvml is None:
+        return None
+
+    is_init = False
+    try:
+        nvml.nvmlInit()
+        is_init = True
+        count = nvml.nvmlDeviceGetCount()
+        if count == 0:
+            return None
+
+        total_gpu_util = 0.0
+        total_temp = 0.0
+        total_gpu_power = 0.0
+        total_power_limit = 0.0
+        total_gpu_max = 0.0
+        for i in range(count):
+            handle = nvml.nvmlDeviceGetHandleByIndex(i)
+            util = nvml.nvmlDeviceGetUtilizationRates(handle)
+            total_gpu_util += float(util.gpu)
+            total_temp += float(
+                nvml.nvmlDeviceGetTemperatureV(handle, nvml.NVML_TEMPERATURE_GPU)
+            )
+            total_gpu_power += float(nvml.nvmlDeviceGetPowerUsage(handle)) / 1000.0
+            with contextlib.suppress(Exception):
+                total_power_limit += (
+                    float(nvml.nvmlDeviceGetPowerManagementLimit(handle)) / 1000.0
+                )
+            with contextlib.suppress(Exception):
+                name = cast(str | bytes, nvml.nvmlDeviceGetName(handle))  # pyright: ignore[reportUnknownMemberType]
+                if isinstance(name, bytes):
+                    name = name.decode("utf-8", errors="replace")
+                total_gpu_max += _device_gpu_max_watts(name)
+
+        # Denominator for the GPU load fraction: prefer NVML's reported power
+        # limit; fall back to the known device GPU-max when NVML reports none
+        # (e.g. GB10, which exposes no power-management limit).
+        power_denominator = (
+            total_power_limit if total_power_limit > 0 else total_gpu_max
+        )
+        gpu_load_fraction = (
+            min(total_gpu_power / power_denominator, 1.0)
+            if power_denominator > 0
+            else 0.0
+        )
+        estimated_cpu_power = (
+            _CPU_POWER_IDLE + (_CPU_POWER_MAX - _CPU_POWER_IDLE) * gpu_load_fraction
+        )
+
+        return NvmlMetrics(
+            system_profile=SystemPerformanceProfile(
+                gpu_usage=total_gpu_util / count / 100.0,
+                temp=total_temp / count,
+                sys_power=total_gpu_power + estimated_cpu_power,
+            ),
+            provides_sys_power=provides_sys_power,
+        )
+    except Exception:
+        return None
+    finally:
+        if is_init:
+            nvml.nvmlShutdown()
diff --git a/src/exo/utils/info_gatherer/spbm.py b/src/exo/utils/info_gatherer/spbm.py
new file mode 100644
index 0000000000..b1dbde518e
--- /dev/null
+++ b/src/exo/utils/info_gatherer/spbm.py
@@ -0,0 +1,89 @@
+import re
+import shutil
+
+import anyio
+from anyio import fail_after
+from loguru import logger
+
+from exo.shared.types.profiling import SystemPerformanceProfile
+from exo.utils.pydantic_ext import TaggedModel
+
+_SPBM_CHIP = "spbm-acpi-0"
+
+_FIELD_RE = re.compile(r"^\s+(\w+):\s+(-?\d+(?:\.\d+)?)\s*$")
+
+
+class SpbmMetrics(TaggedModel):
+    system_profile: SystemPerformanceProfile
+
+
+async def has_spbm() -> bool:
+    if shutil.which("sensors") is None:
+        return False
+    try:
+        with fail_after(5):
+            proc = await anyio.run_process(["sensors", "-u", _SPBM_CHIP], check=False)
+    except (TimeoutError, OSError):
+        return False
+    return proc.returncode == 0 and b"power1_input" in proc.stdout
+
+
+def _parse_sensors_output(output: str) -> dict[str, dict[str, float]]:
+    """Parse `sensors -u` output, keyed by category then channel label.
+
+    Returns {"power": {label: watts}, "temp": {label: celsius},
+             "energy": {label: joules}}. Multiple sections may share a label
+     (e.g. "gpu" appears under power, temp, and energy); categorising by the
+     numeric field's prefix disambiguates them.
+    """
+    result: dict[str, dict[str, float]] = {"power": {}, "temp": {}, "energy": {}}
+    current_label: str | None = None
+    for line in output.splitlines():
+        if not line:
+            continue
+        if line[0] not in " \t":
+            stripped = line.rstrip()
+            current_label = stripped[:-1] if stripped.endswith(":") else None
+            continue
+        match = _FIELD_RE.match(line)
+        if match is None or current_label is None:
+            continue
+        field = match.group(1)
+        try:
+            value = float(match.group(2))
+        except ValueError:
+            continue
+        if not field.endswith("_input"):
+            continue
+        for category in result:
+            if field.startswith(category):
+                result[category][current_label] = value
+                break
+    return result
+
+
+async def gather_spbm_metrics() -> SpbmMetrics | None:
+    try:
+        with fail_after(5):
+            proc = await anyio.run_process(["sensors", "-u", _SPBM_CHIP], check=False)
+    except (TimeoutError, OSError) as e:
+        logger.opt(exception=e).debug("spbm sensors invocation failed")
+        return None
+    if proc.returncode != 0:
+        return None
+
+    parsed = _parse_sensors_output(proc.stdout.decode("utf-8", errors="replace"))
+    sys_total = parsed["power"].get("sys_total")
+    if sys_total is None:
+        return None
+    tj_max = parsed["temp"].get("tj_max")
+    gpu_temp = parsed["temp"].get("gpu")
+
+    temp = tj_max if tj_max is not None else (gpu_temp if gpu_temp is not None else 0.0)
+
+    return SpbmMetrics(
+        system_profile=SystemPerformanceProfile(
+            temp=temp,
+            sys_power=sys_total,
+        ),
+    )
diff --git a/src/exo/utils/info_gatherer/system_info.py b/src/exo/utils/info_gatherer/system_info.py
index 0cc13f9c9c..c4e517695d 100644
--- a/src/exo/utils/info_gatherer/system_info.py
+++ b/src/exo/utils/info_gatherer/system_info.py
@@ -1,6 +1,7 @@
 import platform
 import socket
 import sys
+from pathlib import Path
 from subprocess import CalledProcessError
 
 import psutil
@@ -117,12 +118,90 @@ async def get_network_interfaces() -> list[NetworkInterfaceInfo]:
     return interfaces_info
 
 
+def _read_dmi_field(name: str) -> str | None:
+    try:
+        path = Path(f"/sys/class/dmi/id/{name}")
+        if path.exists():
+            return path.read_text().strip()
+    except (OSError, PermissionError):
+        pass
+    return None
+
+
+async def _get_linux_model_and_chip() -> tuple[str, str]:
+    model = "Linux"
+    chip = "Unknown Chip"
+
+    product_name = _read_dmi_field("product_name")
+    sys_vendor = _read_dmi_field("sys_vendor")
+
+    # DGX Spark: DMI product_name may be "DGX_Spark" or "gx10" variant
+    product_lower = (product_name or "").lower()
+    if product_name and ("dgx" in product_lower or "gx10" in product_lower):
+        model = "DGX Spark"
+        try:
+            process = await run_process(
+                ["nvidia-smi", "--query-gpu=name", "--format=csv,noheader"]
+            )
+            gpu_name = process.stdout.decode().strip().split("\n")[0]
+            chip = gpu_name if gpu_name and gpu_name != "[N/A]" else "NVIDIA GB10"
+        except (CalledProcessError, FileNotFoundError):
+            chip = "NVIDIA GB10"
+        return (model, chip)
+
+    # Other NVIDIA systems (sys_vendor contains "NVIDIA")
+    if sys_vendor and "NVIDIA" in sys_vendor:
+        model = product_name.replace("_", " ") if product_name else "NVIDIA System"
+        try:
+            process = await run_process(
+                ["nvidia-smi", "--query-gpu=name", "--format=csv,noheader"]
+            )
+            gpu_name = process.stdout.decode().strip().split("\n")[0]
+            if gpu_name and gpu_name != "[N/A]":
+                chip = gpu_name
+        except (CalledProcessError, FileNotFoundError):
+            pass
+        return (model, chip)
+
+    # Generic Linux — detect laptop vs desktop via chassis_type
+    # SMBIOS chassis types: 8,9,10,14,31,32 = portable/laptop
+    chassis_type = _read_dmi_field("chassis_type")
+    laptop_chassis_types = {"8", "9", "10", "14", "31", "32"}
+    if chassis_type in laptop_chassis_types:
+        model = "Linux Laptop"
+    elif chassis_type is not None:
+        model = "Linux Desktop"
+
+    # Also check for battery as a fallback laptop indicator
+    if model == "Linux" and Path("/sys/class/power_supply/BAT0").exists():
+        model = "Linux Laptop"
+
+    # Use /proc/cpuinfo for chip
+    cpuinfo_path = Path("/proc/cpuinfo")
+    if cpuinfo_path.exists():
+        try:
+            for line in cpuinfo_path.read_text().splitlines():
+                if line.startswith("model name"):
+                    chip = line.split(":", 1)[1].strip()
+                    break
+        except OSError:
+            pass
+
+    return (model, chip)
+
+
 async def get_model_and_chip() -> tuple[str, str]:
-    """Get Mac system information using system_profiler."""
+    """Get system model and chip information.
+
+    On macOS, uses ``system_profiler``.  On Linux, reads DMI data from
+    sysfs and CPU info from ``/proc/cpuinfo``.
+    """
     model = "Unknown Model"
     chip = "Unknown Chip"
 
-    # TODO: better non mac support
+    if sys.platform == "linux":
+        return await _get_linux_model_and_chip()
+
     if sys.platform != "darwin":
         return (model, chip)