migrate model cards to .toml files

app/EXO/EXO/ContentView.swift

@@ -56,11 +56,6 @@ struct ContentView: View {
     }
 
     private var shouldShowLocalNetworkWarning: Bool {
-        // Show warning if local network is not working and EXO is running.
-        // The checker uses a longer timeout on first launch to allow time for
-        // the permission prompt, so this correctly handles both:
-        // 1. User denied permission on first launch
-        // 2. Permission broke after restart (macOS TCC bug)
         if case .notWorking = localNetworkChecker.status {
             return controller.status != .stopped
         }

app/EXO/EXO/Services/LocalNetworkChecker.swift

@@ -5,8 +5,8 @@ import os.log
 /// Checks if the app's local network permission is actually functional.
 ///
 /// macOS local network permission can appear enabled in System Preferences but not
-/// actually work after a restart. This service uses NWConnection to mDNS multicast
-/// to verify actual connectivity.
+/// actually work after a restart. This service detects this by creating a UDP
+/// connection to the mDNS multicast address (224.0.0.251:5353).
 @MainActor
 final class LocalNetworkChecker: ObservableObject {
     enum Status: Equatable {
@@ -35,43 +35,30 @@ final class LocalNetworkChecker: ObservableObject {
     }
 
     private static let logger = Logger(subsystem: "io.exo.EXO", category: "LocalNetworkChecker")
-    private static let hasCompletedInitialCheckKey = "LocalNetworkChecker.hasCompletedInitialCheck"
 
     @Published private(set) var status: Status = .unknown
+    @Published private(set) var lastConnectionState: String = "none"
 
     private var connection: NWConnection?
     private var checkTask: Task<Void, Never>?
 
-    /// Whether we've completed at least one check (stored in UserDefaults)
-    private var hasCompletedInitialCheck: Bool {
-        get { UserDefaults.standard.bool(forKey: Self.hasCompletedInitialCheckKey) }
-        set { UserDefaults.standard.set(newValue, forKey: Self.hasCompletedInitialCheckKey) }
-    }
-
     /// Checks if local network access is working.
     func check() {
         checkTask?.cancel()
         status = .checking
-
-        // Use longer timeout on first launch to allow time for permission prompt
-        let isFirstCheck = !hasCompletedInitialCheck
-        let timeout: UInt64 = isFirstCheck ? 30_000_000_000 : 3_000_000_000
+        lastConnectionState = "connecting"
 
         checkTask = Task { [weak self] in
             guard let self else { return }
-
-            Self.logger.info("Checking local network connectivity (first check: \(isFirstCheck))")
-            let result = await self.checkConnectivity(timeout: timeout)
+            let result = await self.performCheck()
             self.status = result
-            self.hasCompletedInitialCheck = true
-
             Self.logger.info("Local network check complete: \(result.displayText)")
         }
     }
 
-    /// Checks connectivity using NWConnection to mDNS multicast.
-    /// The connection attempt triggers the permission prompt if not yet shown.
-    private func checkConnectivity(timeout: UInt64) async -> Status {
+    private func performCheck() async -> Status {
+        Self.logger.info("Checking local network access via UDP multicast")
+
         connection?.cancel()
         connection = nil
 
@@ -97,7 +84,22 @@ final class LocalNetworkChecker: ObservableObject {
                 continuation.resume(returning: status)
             }
 
-            conn.stateUpdateHandler = { state in
+            conn.stateUpdateHandler = { [weak self] state in
+                let stateStr: String
+                switch state {
+                case .setup: stateStr = "setup"
+                case .preparing: stateStr = "preparing"
+                case .ready: stateStr = "ready"
+                case .waiting(let e): stateStr = "waiting(\(e))"
+                case .failed(let e): stateStr = "failed(\(e))"
+                case .cancelled: stateStr = "cancelled"
+                @unknown default: stateStr = "unknown"
+                }
+
+                Task { @MainActor in
+                    self?.lastConnectionState = stateStr
+                }
+
                 switch state {
                 case .ready:
                     resumeOnce(.working)
@@ -106,7 +108,6 @@ final class LocalNetworkChecker: ObservableObject {
                     if errorStr.contains("54") || errorStr.contains("ECONNRESET") {
                         resumeOnce(.notWorking(reason: "Connection blocked"))
                     }
-                // Otherwise keep waiting - might be showing permission prompt
                 case .failed(let error):
                     let errorStr = "\(error)"
                     if errorStr.contains("65") || errorStr.contains("EHOSTUNREACH")
@@ -126,7 +127,7 @@ final class LocalNetworkChecker: ObservableObject {
             conn.start(queue: .main)
 
             Task {
-                try? await Task.sleep(nanoseconds: timeout)
+                try? await Task.sleep(nanoseconds: 3_000_000_000)
                 let state = conn.state
                 switch state {
                 case .ready:

bench/exo_bench.py

@@ -241,9 +241,6 @@ class PromptSizer:
             ids = tokenizer.apply_chat_template(
                 messages, tokenize=True, add_generation_prompt=True
             )
-            # Fix for transformers 5.x
-            if hasattr(ids, "input_ids"):
-                ids = ids.input_ids
             return int(len(ids))
 
         return count_fn

dashboard/src/lib/components/ChatForm.svelte

@@ -60,39 +60,12 @@
 		return models;
 	});
 
-	// Track previous model IDs to detect newly added models (plain variable to avoid reactive loop)
-	let previousModelIds: Set<string> = new Set();
-
-	// Auto-select the first available model if none is selected, if current selection is stale, or if a new model is added
+	// Auto-select the first available model if none is selected
 	$effect(() => {
 		const models = availableModels();
-		const currentModelIds = new Set(models.map(m => m.id));
-
-		if (models.length > 0) {
-			// Find newly added models (in current but not in previous)
-			const newModels = models.filter(m => !previousModelIds.has(m.id));
-
-			// If no model selected, select the first available
-			if (!currentModel) {
-				setSelectedChatModel(models[0].id);
-			}
-			// If current model is stale (no longer has a running instance), reset to first available
-			else if (!models.some(m => m.id === currentModel)) {
-				setSelectedChatModel(models[0].id);
-			}
-			// If a new model was just added, select it
-			else if (newModels.length > 0 && previousModelIds.size > 0) {
-				setSelectedChatModel(newModels[0].id);
-			}
-		} else {
-			// No instances running - clear the selected model
-			if (currentModel) {
-				setSelectedChatModel('');
-			}
+		if (models.length > 0 && !currentModel) {
+			setSelectedChatModel(models[0].id);
 		}
-
-		// Update previous model IDs for next comparison
-		previousModelIds = currentModelIds;
 	});
 
 	function getInstanceModelId(instanceWrapped: unknown): string {

dashboard/src/routes/+page.svelte

@@ -400,8 +400,10 @@ function toggleInstanceDownloadDetails(nodeId: string): void {
 				const errorText = await response.text();
 				console.error('Failed to launch instance:', errorText);
 			} else {
-				// Always auto-select the newly launched model so the user chats to what they just launched
-				setSelectedChatModel(modelId);
+				// Auto-select the launched model only if no model is currently selected
+				if (!selectedChatModel()) {
+					setSelectedChatModel(modelId);
+				}
 				
 				// Scroll to the bottom of instances container to show the new instance
 				// Use multiple attempts to ensure DOM has updated with the new instance
@@ -761,10 +763,6 @@ function toggleInstanceDownloadDetails(nodeId: string): void {
 	async function deleteInstance(instanceId: string) {
 		if (!confirm(`Delete instance ${instanceId.slice(0, 8)}...?`)) return;
 		
-		// Get the model ID of the instance being deleted before we delete it
-		const deletedInstanceModelId = getInstanceModelId(instanceData[instanceId]);
-		const wasSelected = selectedChatModel() === deletedInstanceModelId;
-		
 		try {
 			const response = await fetch(`/instance/${instanceId}`, {
 				method: 'DELETE',
@@ -773,24 +771,6 @@ function toggleInstanceDownloadDetails(nodeId: string): void {
 			
 			if (!response.ok) {
 				console.error('Failed to delete instance:', response.status);
-			} else if (wasSelected) {
-				// If we deleted the currently selected model, switch to another available model
-				// Find another instance that isn't the one we just deleted
-				const remainingInstances = Object.entries(instanceData).filter(([id]) => id !== instanceId);
-				if (remainingInstances.length > 0) {
-					// Select the last instance (most recently added, since objects preserve insertion order)
-					const [, lastInstance] = remainingInstances[remainingInstances.length - 1];
-					const newModelId = getInstanceModelId(lastInstance);
-					if (newModelId && newModelId !== 'Unknown' && newModelId !== 'Unknown Model') {
-						setSelectedChatModel(newModelId);
-					} else {
-						// Clear selection if no valid model found
-						setSelectedChatModel('');
-					}
-				} else {
-					// No more instances, clear the selection
-					setSelectedChatModel('');
-				}
 			}
 		} catch (error) {
 			console.error('Error deleting instance:', error);

justfile

@@ -1,5 +1,3 @@
-export NIX_CONFIG := "extra-experimental-features = nix-command flakes"
-
 fmt:
     nix fmt
 

pyproject.toml

@@ -23,7 +23,7 @@ dependencies = [
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",
     "openai-harmony>=0.0.8",
-    "httpx>=0.28.1",
+    "tomlkit>=0.14.0",
 ]
 
 [project.scripts]

resources/model_cards/deepseek-v3.1-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "deepseek-v3.1-4bit"
+model_id = "mlx-community/DeepSeek-V3.1-4bit"
+name = "DeepSeek V3.1 (4-bit)"
+description = "DeepSeek V3.1 is a large language model trained on the DeepSeek V3.1 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/DeepSeek-V3.1-4bit"
+pretty_name = "DeepSeek V3.1 (4-bit)"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 405874409472

resources/model_cards/deepseek-v3.1-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "deepseek-v3.1-8bit"
+model_id = "mlx-community/DeepSeek-V3.1-8bit"
+name = "DeepSeek V3.1 (8-bit)"
+description = "DeepSeek V3.1 is a large language model trained on the DeepSeek V3.1 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/DeepSeek-V3.1-8bit"
+pretty_name = "DeepSeek V3.1 (8-bit)"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 765577920512

resources/model_cards/glm-4.5-air-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "glm-4.5-air-8bit"
+model_id = "mlx-community/GLM-4.5-Air-8bit"
+name = "GLM 4.5 Air 8bit"
+description = "GLM 4.5 Air 8bit"
+tags = []
+
+[metadata]
+model_id = "mlx-community/GLM-4.5-Air-8bit"
+pretty_name = "GLM 4.5 Air 8bit"
+n_layers = 46
+hidden_size = 4096
+supports_tensor = false
+
+[metadata.storage_size]
+in_bytes = 122406567936

resources/model_cards/glm-4.5-air-bf16.toml

@@ -0,0 +1,15 @@
+short_id = "glm-4.5-air-bf16"
+model_id = "mlx-community/GLM-4.5-Air-bf16"
+name = "GLM 4.5 Air bf16"
+description = "GLM 4.5 Air bf16"
+tags = []
+
+[metadata]
+model_id = "mlx-community/GLM-4.5-Air-bf16"
+pretty_name = "GLM 4.5 Air bf16"
+n_layers = 46
+hidden_size = 4096
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 229780750336

resources/model_cards/glm-4.7-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "glm-4.7-4bit"
+model_id = "mlx-community/GLM-4.7-4bit"
+name = "GLM 4.7 4bit"
+description = "GLM 4.7 4bit"
+tags = []
+
+[metadata]
+model_id = "mlx-community/GLM-4.7-4bit"
+pretty_name = "GLM 4.7 4bit"
+n_layers = 91
+hidden_size = 5120
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 198556925568

resources/model_cards/glm-4.7-6bit.toml

@@ -0,0 +1,15 @@
+short_id = "glm-4.7-6bit"
+model_id = "mlx-community/GLM-4.7-6bit"
+name = "GLM 4.7 6bit"
+description = "GLM 4.7 6bit"
+tags = []
+
+[metadata]
+model_id = "mlx-community/GLM-4.7-6bit"
+pretty_name = "GLM 4.7 6bit"
+n_layers = 91
+hidden_size = 5120
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 286737579648

resources/model_cards/glm-4.7-8bit-gs32.toml

@@ -0,0 +1,15 @@
+short_id = "glm-4.7-8bit-gs32"
+model_id = "mlx-community/GLM-4.7-8bit-gs32"
+name = "GLM 4.7 8bit (gs32)"
+description = "GLM 4.7 8bit (gs32)"
+tags = []
+
+[metadata]
+model_id = "mlx-community/GLM-4.7-8bit-gs32"
+pretty_name = "GLM 4.7 8bit (gs32)"
+n_layers = 91
+hidden_size = 5120
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 396963397248

resources/model_cards/gpt-oss-120b-MXFP4-Q8.toml

@@ -0,0 +1,15 @@
+short_id = "gpt-oss-120b-MXFP4-Q8"
+model_id = "mlx-community/gpt-oss-120b-MXFP4-Q8"
+name = "GPT-OSS 120B (MXFP4-Q8, MLX)"
+description = "OpenAI's GPT-OSS 120B is a 117B-parameter Mixture-of-Experts model designed for high-reasoning and general-purpose use; this variant is a 4-bit MLX conversion for Apple Silicon."
+tags = []
+
+[metadata]
+model_id = "mlx-community/gpt-oss-120b-MXFP4-Q8"
+pretty_name = "GPT-OSS 120B (MXFP4-Q8, MLX)"
+n_layers = 36
+hidden_size = 2880
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 70652212224

resources/model_cards/gpt-oss-20b-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "gpt-oss-20b-4bit"
+model_id = "mlx-community/gpt-oss-20b-MXFP4-Q4"
+name = "GPT-OSS 20B (MXFP4-Q4, MLX)"
+description = "OpenAI's GPT-OSS 20B is a medium-sized MoE model for lower-latency and local or specialized use cases; this MLX variant uses MXFP4 4-bit quantization."
+tags = []
+
+[metadata]
+model_id = "mlx-community/gpt-oss-20b-MXFP4-Q4"
+pretty_name = "GPT-OSS 20B (MXFP4-Q4, MLX)"
+n_layers = 24
+hidden_size = 2880
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 12025908224

resources/model_cards/kimi-k2-instruct-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "kimi-k2-instruct-4bit"
+model_id = "mlx-community/Kimi-K2-Instruct-4bit"
+name = "Kimi K2 Instruct (4-bit)"
+description = "Kimi K2 is a large language model trained on the Kimi K2 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Kimi-K2-Instruct-4bit"
+pretty_name = "Kimi K2 Instruct (4-bit)"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 620622774272

resources/model_cards/kimi-k2-thinking.toml

@@ -0,0 +1,15 @@
+short_id = "kimi-k2-thinking"
+model_id = "mlx-community/Kimi-K2-Thinking"
+name = "Kimi K2 Thinking (4-bit)"
+description = "Kimi K2 Thinking is the latest, most capable version of open-source thinking model."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Kimi-K2-Thinking"
+pretty_name = "Kimi K2 Thinking (4-bit)"
+n_layers = 61
+hidden_size = 7168
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 706522120192

resources/model_cards/llama-3.1-70b.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.1-70b"
+model_id = "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"
+name = "Llama 3.1 70B (4-bit)"
+description = "Llama 3.1 is a large language model trained on the Llama 3.1 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"
+pretty_name = "Llama 3.1 70B (4-bit)"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 40652242944

resources/model_cards/llama-3.1-8b-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.1-8b-8bit"
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"
+name = "Llama 3.1 8B (8-bit)"
+description = "Llama 3.1 is a large language model trained on the Llama 3.1 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"
+pretty_name = "Llama 3.1 8B (8-bit)"
+n_layers = 32
+hidden_size = 4096
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 8954839040

resources/model_cards/llama-3.1-8b-bf16.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.1-8b-bf16"
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"
+name = "Llama 3.1 8B (BF16)"
+description = "Llama 3.1 is a large language model trained on the Llama 3.1 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"
+pretty_name = "Llama 3.1 8B (BF16)"
+n_layers = 32
+hidden_size = 4096
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 16882073600

resources/model_cards/llama-3.1-8b.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.1-8b"
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
+name = "Llama 3.1 8B (4-bit)"
+description = "Llama 3.1 is a large language model trained on the Llama 3.1 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
+pretty_name = "Llama 3.1 8B (4-bit)"
+n_layers = 32
+hidden_size = 4096
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 4637851648

resources/model_cards/llama-3.2-1b.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.2-1b"
+model_id = "mlx-community/Llama-3.2-1B-Instruct-4bit"
+name = "Llama 3.2 1B (4-bit)"
+description = "Llama 3.2 is a large language model trained on the Llama 3.2 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Llama-3.2-1B-Instruct-4bit"
+pretty_name = "Llama 3.2 1B (4-bit)"
+n_layers = 16
+hidden_size = 2048
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 729808896

resources/model_cards/llama-3.2-3b-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.2-3b-8bit"
+model_id = "mlx-community/Llama-3.2-3B-Instruct-8bit"
+name = "Llama 3.2 3B (8-bit)"
+description = "Llama 3.2 is a large language model trained on the Llama 3.2 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Llama-3.2-3B-Instruct-8bit"
+pretty_name = "Llama 3.2 3B (8-bit)"
+n_layers = 28
+hidden_size = 3072
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 3501195264

resources/model_cards/llama-3.2-3b.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.2-3b"
+model_id = "mlx-community/Llama-3.2-3B-Instruct-4bit"
+name = "Llama 3.2 3B (4-bit)"
+description = "Llama 3.2 is a large language model trained on the Llama 3.2 dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Llama-3.2-3B-Instruct-4bit"
+pretty_name = "Llama 3.2 3B (4-bit)"
+n_layers = 28
+hidden_size = 3072
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 1863319552

resources/model_cards/llama-3.3-70b-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.3-70b-8bit"
+model_id = "mlx-community/Llama-3.3-70B-Instruct-8bit"
+name = "Llama 3.3 70B (8-bit)"
+description = "The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out)"
+tags = []
+
+[metadata]
+model_id = "mlx-community/Llama-3.3-70B-Instruct-8bit"
+pretty_name = "Llama 3.3 70B (8-bit)"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 76799803392

resources/model_cards/llama-3.3-70b-fp16.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.3-70b-fp16"
+model_id = "mlx-community/llama-3.3-70b-instruct-fp16"
+name = "Llama 3.3 70B (FP16)"
+description = "The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out)"
+tags = []
+
+[metadata]
+model_id = "mlx-community/llama-3.3-70b-instruct-fp16"
+pretty_name = "Llama 3.3 70B (FP16)"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 144383672320

resources/model_cards/llama-3.3-70b.toml

@@ -0,0 +1,15 @@
+short_id = "llama-3.3-70b"
+model_id = "mlx-community/Llama-3.3-70B-Instruct-4bit"
+name = "Llama 3.3 70B (4-bit)"
+description = "The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out)"
+tags = []
+
+[metadata]
+model_id = "mlx-community/Llama-3.3-70B-Instruct-4bit"
+pretty_name = "Llama 3.3 70B"
+n_layers = 80
+hidden_size = 8192
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 40652242944

resources/model_cards/minimax-m2.1-3bit.toml

@@ -0,0 +1,15 @@
+short_id = "minimax-m2.1-3bit"
+model_id = "mlx-community/MiniMax-M2.1-3bit"
+name = "MiniMax M2.1 3bit"
+description = "MiniMax M2.1 3bit"
+tags = []
+
+[metadata]
+model_id = "mlx-community/MiniMax-M2.1-3bit"
+pretty_name = "MiniMax M2.1 3bit"
+n_layers = 61
+hidden_size = 3072
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 100086644736

resources/model_cards/minimax-m2.1-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "minimax-m2.1-8bit"
+model_id = "mlx-community/MiniMax-M2.1-8bit"
+name = "MiniMax M2.1 8bit"
+description = "MiniMax M2.1 8bit"
+tags = []
+
+[metadata]
+model_id = "mlx-community/MiniMax-M2.1-8bit"
+pretty_name = "MiniMax M2.1 8bit"
+n_layers = 61
+hidden_size = 3072
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 242986745856

resources/model_cards/qwen3-0.6b-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-0.6b-8bit"
+model_id = "mlx-community/Qwen3-0.6B-8bit"
+name = "Qwen3 0.6B (8-bit)"
+description = "Qwen3 0.6B is a large language model trained on the Qwen3 0.6B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-0.6B-8bit"
+pretty_name = "Qwen3 0.6B (8-bit)"
+n_layers = 28
+hidden_size = 1024
+supports_tensor = false
+
+[metadata.storage_size]
+in_bytes = 698351616

resources/model_cards/qwen3-0.6b.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-0.6b"
+model_id = "mlx-community/Qwen3-0.6B-4bit"
+name = "Qwen3 0.6B (4-bit)"
+description = "Qwen3 0.6B is a large language model trained on the Qwen3 0.6B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-0.6B-4bit"
+pretty_name = "Qwen3 0.6B (4-bit)"
+n_layers = 28
+hidden_size = 1024
+supports_tensor = false
+
+[metadata.storage_size]
+in_bytes = 342884352

resources/model_cards/qwen3-235b-a22b-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-235b-a22b-4bit"
+model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"
+name = "Qwen3 235B A22B (4-bit)"
+description = "Qwen3 235B (Active 22B) is a large language model trained on the Qwen3 235B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"
+pretty_name = "Qwen3 235B A22B (4-bit)"
+n_layers = 94
+hidden_size = 4096
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 141733920768

resources/model_cards/qwen3-235b-a22b-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-235b-a22b-8bit"
+model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"
+name = "Qwen3 235B A22B (8-bit)"
+description = "Qwen3 235B (Active 22B) is a large language model trained on the Qwen3 235B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"
+pretty_name = "Qwen3 235B A22B (8-bit)"
+n_layers = 94
+hidden_size = 4096
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 268435456000

resources/model_cards/qwen3-30b-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-30b-8bit"
+model_id = "mlx-community/Qwen3-30B-A3B-8bit"
+name = "Qwen3 30B A3B (8-bit)"
+description = "Qwen3 30B is a large language model trained on the Qwen3 30B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-30B-A3B-8bit"
+pretty_name = "Qwen3 30B A3B (8-bit)"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 33279705088

resources/model_cards/qwen3-30b.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-30b"
+model_id = "mlx-community/Qwen3-30B-A3B-4bit"
+name = "Qwen3 30B A3B (4-bit)"
+description = "Qwen3 30B is a large language model trained on the Qwen3 30B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-30B-A3B-4bit"
+pretty_name = "Qwen3 30B A3B (4-bit)"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 17612931072

resources/model_cards/qwen3-80b-a3B-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-80b-a3B-4bit"
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"
+name = "Qwen3 80B A3B (4-bit)"
+description = "Qwen3 80B"
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"
+pretty_name = "Qwen3 80B A3B (4-bit)"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 46976204800

resources/model_cards/qwen3-80b-a3B-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-80b-a3B-8bit"
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"
+name = "Qwen3 80B A3B (8-bit)"
+description = "Qwen3 80B"
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"
+pretty_name = "Qwen3 80B A3B (8-bit)"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 88814387200

resources/model_cards/qwen3-80b-a3B-thinking-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-80b-a3B-thinking-4bit"
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"
+name = "Qwen3 80B A3B Thinking (4-bit)"
+description = "Qwen3 80B Reasoning model"
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"
+pretty_name = "Qwen3 80B A3B (4-bit)"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 88814387200

resources/model_cards/qwen3-80b-a3B-thinking-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-80b-a3B-thinking-8bit"
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"
+name = "Qwen3 80B A3B Thinking (8-bit)"
+description = "Qwen3 80B Reasoning model"
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"
+pretty_name = "Qwen3 80B A3B (8-bit)"
+n_layers = 48
+hidden_size = 2048
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 88814387200

resources/model_cards/qwen3-coder-480b-a35b-4bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-coder-480b-a35b-4bit"
+model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"
+name = "Qwen3 Coder 480B A35B (4-bit)"
+description = "Qwen3 Coder 480B (Active 35B) is a large language model trained on the Qwen3 Coder 480B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"
+pretty_name = "Qwen3 Coder 480B A35B (4-bit)"
+n_layers = 62
+hidden_size = 6144
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 289910292480

resources/model_cards/qwen3-coder-480b-a35b-8bit.toml

@@ -0,0 +1,15 @@
+short_id = "qwen3-coder-480b-a35b-8bit"
+model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"
+name = "Qwen3 Coder 480B A35B (8-bit)"
+description = "Qwen3 Coder 480B (Active 35B) is a large language model trained on the Qwen3 Coder 480B dataset."
+tags = []
+
+[metadata]
+model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"
+pretty_name = "Qwen3 Coder 480B A35B (8-bit)"
+n_layers = 62
+hidden_size = 6144
+supports_tensor = true
+
+[metadata.storage_size]
+in_bytes = 579820584960

src/exo/master/api.py

@@ -13,6 +13,12 @@ from hypercorn.asyncio import serve  # pyright: ignore[reportUnknownVariableType
 from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
 from loguru import logger
+from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
+    HarmonyEncodingName,
+    Role,
+    StreamableParser,
+    load_harmony_encoding,
+)
 
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
@@ -61,6 +67,8 @@ from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.dashboard_path import find_dashboard
 from exo.utils.event_buffer import OrderedBuffer
 
+encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
+
 
 def chunk_to_response(
     chunk: TokenChunk, command_id: CommandId
@@ -373,8 +381,35 @@ class API:
             instance_id=instance_id,
         )
 
+    async def _process_gpt_oss(self, token_chunks: Receiver[TokenChunk]):
+        stream = StreamableParser(encoding, role=Role.ASSISTANT)
+        thinking = False
+
+        async for chunk in token_chunks:
+            stream.process(chunk.token_id)
+
+            delta = stream.last_content_delta
+            ch = stream.current_channel
+
+            if ch == "analysis" and not thinking:
+                thinking = True
+                yield chunk.model_copy(update={"text": "<think>"})
+
+            if ch != "analysis" and thinking:
+                thinking = False
+                yield chunk.model_copy(update={"text": "</think>"})
+
+            if delta:
+                yield chunk.model_copy(update={"text": delta})
+
+            if chunk.finish_reason is not None:
+                if thinking:
+                    yield chunk.model_copy(update={"text": "</think>"})
+                yield chunk
+                break
+
     async def _chat_chunk_stream(
-        self, command_id: CommandId
+        self, command_id: CommandId, parse_gpt_oss: bool
     ) -> AsyncGenerator[TokenChunk, None]:
         """Yield `TokenChunk`s for a given command until completion."""
 
@@ -382,10 +417,16 @@ class API:
             self._chat_completion_queues[command_id], recv = channel[TokenChunk]()
 
             with recv as token_chunks:
-                async for chunk in token_chunks:
-                    yield chunk
-                    if chunk.finish_reason is not None:
-                        break
+                if parse_gpt_oss:
+                    async for chunk in self._process_gpt_oss(token_chunks):
+                        yield chunk
+                        if chunk.finish_reason is not None:
+                            break
+                else:
+                    async for chunk in token_chunks:
+                        yield chunk
+                        if chunk.finish_reason is not None:
+                            break
 
         except anyio.get_cancelled_exc_class():
             # TODO: TaskCancelled
@@ -401,11 +442,11 @@ class API:
             del self._chat_completion_queues[command_id]
 
     async def _generate_chat_stream(
-        self, command_id: CommandId
+        self, command_id: CommandId, parse_gpt_oss: bool
     ) -> AsyncGenerator[str, None]:
         """Generate chat completion stream as JSON strings."""
 
-        async for chunk in self._chat_chunk_stream(command_id):
+        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
             chunk_response: ChatCompletionResponse = chunk_to_response(
                 chunk, command_id
             )
@@ -417,7 +458,7 @@ class API:
                 yield "data: [DONE]\n\n"
 
     async def _collect_chat_completion(
-        self, command_id: CommandId
+        self, command_id: CommandId, parse_gpt_oss: bool
     ) -> ChatCompletionResponse:
         """Collect all token chunks for a chat completion and return a single response."""
 
@@ -425,7 +466,7 @@ class API:
         model: str | None = None
         finish_reason: FinishReason | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id):
+        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
             if model is None:
                 model = chunk.model
 
@@ -454,7 +495,7 @@ class API:
         )
 
     async def _collect_chat_completion_with_stats(
-        self, command_id: CommandId
+        self, command_id: CommandId, parse_gpt_oss: bool
     ) -> BenchChatCompletionResponse:
         text_parts: list[str] = []
         model: str | None = None
@@ -462,7 +503,7 @@ class API:
 
         stats: GenerationStats | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id):
+        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
             if model is None:
                 model = chunk.model
 
@@ -503,6 +544,8 @@ class API:
         """Handle chat completions, supporting both streaming and non-streaming responses."""
         model_meta = await resolve_model_meta(payload.model)
         payload.model = model_meta.model_id
+        parse_gpt_oss = "gpt-oss" in model_meta.model_id.lower()
+        logger.info(f"{parse_gpt_oss=}")
 
         if not any(
             instance.shard_assignments.model_id == payload.model
@@ -519,16 +562,17 @@ class API:
         await self._send(command)
         if payload.stream:
             return StreamingResponse(
-                self._generate_chat_stream(command.command_id),
+                self._generate_chat_stream(command.command_id, parse_gpt_oss),
                 media_type="text/event-stream",
             )
 
-        return await self._collect_chat_completion(command.command_id)
+        return await self._collect_chat_completion(command.command_id, parse_gpt_oss)
 
     async def bench_chat_completions(
         self, payload: BenchChatCompletionTaskParams
     ) -> BenchChatCompletionResponse:
         model_meta = await resolve_model_meta(payload.model)
+        parse_gpt_oss = "gpt-oss" in model_meta.model_id.lower()
         payload.model = model_meta.model_id
 
         if not any(
@@ -545,7 +589,10 @@ class API:
         command = ChatCompletion(request_params=payload)
         await self._send(command)
 
-        response = await self._collect_chat_completion_with_stats(command.command_id)
+        response = await self._collect_chat_completion_with_stats(
+            command.command_id,
+            parse_gpt_oss,
+        )
         return response
 
     def _calculate_total_available_memory(self) -> Memory:

src/exo/master/placement_utils.py

@@ -49,22 +49,20 @@ def get_smallest_cycles(cycles: list[list[NodeInfo]]) -> list[list[NodeInfo]]:
     return [cycle for cycle in cycles if len(cycle) == min_nodes]
 
 
-def _assign_layers_by_ram(
+def get_shard_assignments_for_pipeline_parallel(
     model_meta: ModelMetadata,
     selected_cycle: list[NodeWithProfile],
-) -> ShardAssignments:
-    """Assign layers proportionally based on available RAM."""
+):
+    cycle_memory = sum(
+        (node.node_profile.memory.ram_available for node in selected_cycle),
+        start=Memory(),
+    )
     total_layers = model_meta.n_layers
     world_size = len(selected_cycle)
     runner_to_shard: dict[RunnerId, ShardMetadata] = {}
     node_to_runner: dict[NodeId, RunnerId] = {}
 
-    cycle_memory = sum(
-        (node.node_profile.memory.ram_available for node in selected_cycle),
-        start=Memory(),
-    )
     layers_assigned = 0
-
     for i, node in enumerate(selected_cycle):
         if i == len(selected_cycle) - 1:
             node_layers = total_layers - layers_assigned
@@ -79,6 +77,7 @@ def _assign_layers_by_ram(
             node_layers = max(1, node_layers)
 
         runner_id = RunnerId()
+
         shard = PipelineShardMetadata(
             model_meta=model_meta,
             device_rank=i,
@@ -87,143 +86,18 @@ def _assign_layers_by_ram(
             end_layer=layers_assigned + node_layers,
             n_layers=total_layers,
         )
+
         runner_to_shard[runner_id] = shard
         node_to_runner[node.node_id] = runner_id
         layers_assigned += node_layers
 
-    return ShardAssignments(
+    shard_assignments = ShardAssignments(
         model_id=model_meta.model_id,
         runner_to_shard=runner_to_shard,
         node_to_runner=node_to_runner,
     )
 
-
-def _reserve_base_layers(world_size: int, total_layers: int) -> dict[int, int]:
-    """Reserve 1 layer per node to ensure connectivity."""
-    assignments = {i: 0 for i in range(world_size)}
-    remaining_layers = total_layers
-
-    for i in range(world_size):
-        assignments[i] = 1
-        remaining_layers -= 1
-
-    if remaining_layers < 0:
-        logger.warning(
-            "Fewer layers than nodes! Reducing to 1 layer per node where possible."
-        )
-        assignments = {i: 1 if i < total_layers else 0 for i in range(world_size)}
-        remaining_layers = 0
-
-    return assignments
-
-
-def _distribute_layers_by_bandwidth(
-    selected_cycle: list[NodeWithProfile],
-    assignments: dict[int, int],
-    remaining_layers: int,
-    model_meta: ModelMetadata,
-) -> None:
-    """Distribute remaining layers based on bandwidth and RAM capacity."""
-    indexed_nodes = list(enumerate(selected_cycle))
-    sorted_nodes = sorted(
-        indexed_nodes,
-        key=lambda x: x[1].node_profile.memory_bandwidth or 0,
-        reverse=True,
-    )
-
-    for original_idx, node in sorted_nodes:
-        if remaining_layers <= 0:
-            break
-
-        layer_size_bytes = model_meta.storage_size.in_bytes / model_meta.n_layers
-        max_layers_by_ram = int(
-            node.node_profile.memory.ram_available.in_bytes // layer_size_bytes
-        )
-        can_take = max(0, max_layers_by_ram - assignments[original_idx])
-        take = min(can_take, remaining_layers)
-        assignments[original_idx] += take
-        remaining_layers -= take
-
-    if remaining_layers > 0:
-        logger.warning(
-            "All nodes maxed out on RAM estimation, dumping remaining layers on fastest nodes."
-        )
-        for original_idx, _ in sorted_nodes:
-            assignments[original_idx] += 1
-            remaining_layers -= 1
-            if remaining_layers == 0:
-                break
-
-
-def _create_shard_assignments(
-    model_meta: ModelMetadata,
-    selected_cycle: list[NodeWithProfile],
-    assignments: dict[int, int],
-) -> ShardAssignments:
-    """Create shard assignments from layer assignments."""
-    world_size = len(selected_cycle)
-    runner_to_shard: dict[RunnerId, ShardMetadata] = {}
-    node_to_runner: dict[NodeId, RunnerId] = {}
-
-    current_start = 0
-    for i, node in enumerate(selected_cycle):
-        count = assignments[i]
-        runner_id = RunnerId()
-        shard = PipelineShardMetadata(
-            model_meta=model_meta,
-            device_rank=i,
-            world_size=world_size,
-            start_layer=current_start,
-            end_layer=current_start + count,
-            n_layers=model_meta.n_layers,
-        )
-        runner_to_shard[runner_id] = shard
-        node_to_runner[node.node_id] = runner_id
-        current_start += count
-
-    return ShardAssignments(
-        model_id=model_meta.model_id,
-        runner_to_shard=runner_to_shard,
-        node_to_runner=node_to_runner,
-    )
-
-
-def _assign_layers_by_bandwidth(
-    model_meta: ModelMetadata,
-    selected_cycle: list[NodeWithProfile],
-) -> ShardAssignments:
-    """Assign layers based on memory bandwidth."""
-    logger.info("Using bandwidth-aware shard assignment")
-
-    total_layers = model_meta.n_layers
-    world_size = len(selected_cycle)
-
-    assignments = _reserve_base_layers(world_size, total_layers)
-    remaining_layers = total_layers - sum(assignments.values())
-
-    if remaining_layers > 0:
-        _distribute_layers_by_bandwidth(
-            selected_cycle, assignments, remaining_layers, model_meta
-        )
-
-    return _create_shard_assignments(model_meta, selected_cycle, assignments)
-
-
-def get_shard_assignments_for_pipeline_parallel(
-    model_meta: ModelMetadata,
-    selected_cycle: list[NodeWithProfile],
-):
-    has_bandwidth = all(
-        node.node_profile.memory_bandwidth is not None for node in selected_cycle
-    )
-
-    if not has_bandwidth:
-        logger.info(
-            "Bandwidth data missing for some nodes, falling back to RAM-proportional assignment"
-        )
-        return _assign_layers_by_ram(model_meta, selected_cycle)
-
-    return _assign_layers_by_bandwidth(model_meta, selected_cycle)
+    return shard_assignments
 
 
 def get_shard_assignments_for_tensor_parallel(

src/exo/master/tests/test_placement_utils.py

@@ -397,106 +397,3 @@ def test_get_mlx_jaccl_coordinators(
     assert coordinators[node_c_id] == (
         f"{conn_c_a.send_back_multiaddr.ip_address}:5000"
     ), "node_c should use the IP from conn_c_a"
-
-
-def test_get_shard_assignments_bandwidth_aware(
-    topology: Topology,
-    create_node: Callable[[int, NodeId | None], NodeInfo],
-    create_connection: Callable[[NodeId, NodeId], Connection],
-):
-    # arrange
-    node_a_id = NodeId()
-    node_b_id = NodeId()
-    node_c_id = NodeId()
-
-    # Create nodes with identical RAM (plenty of it)
-    # Using 1GB to ensure no RAM constraints (model is small)
-    node_a = create_node(1024 * 1024 * 1024, node_a_id)
-    node_b = create_node(1024 * 1024 * 1024, node_b_id)
-    node_c = create_node(1024 * 1024 * 1024, node_c_id)
-
-    # Set Bandwidths: A=400 (Fastest), B=200, C=100 (Slowest)
-    assert node_a.node_profile is not None
-    assert node_b.node_profile is not None
-    assert node_c.node_profile is not None
-
-    node_a.node_profile.memory_bandwidth = 400_000_000_000
-    node_b.node_profile.memory_bandwidth = 200_000_000_000
-    node_c.node_profile.memory_bandwidth = 100_000_000_000
-
-    topology.add_node(node_a)
-    topology.add_node(node_b)
-    topology.add_node(node_c)
-
-    topology.add_connection(create_connection(node_a_id, node_b_id))
-    topology.add_connection(create_connection(node_b_id, node_c_id))
-    topology.add_connection(create_connection(node_c_id, node_a_id))
-
-    # Needs full cycle edges for get_cycles/get_shard_assignments if strict?
-    # Actually get_cycles just looks for cycles.
-    # But let's follow the pattern of other tests if they add bidirectional.
-    # checking test_filter_cycles_by_memory, it adds both directions.
-    topology.add_connection(create_connection(node_b_id, node_a_id))
-    topology.add_connection(create_connection(node_c_id, node_b_id))
-    topology.add_connection(create_connection(node_a_id, node_c_id))
-
-    model_meta = ModelMetadata(
-        model_id=ModelId("test-model"),
-        pretty_name="Test Model",
-        n_layers=30,  # 30 layers
-        storage_size=Memory.from_kb(
-            300
-        ),  # 10KB per layer. Nodes have 100MB RAM (100*1024 in create_node usually means KB? other tests use 1000*1024).
-        # create_node arg is likely KB or Bytes.
-        # test_filter_cycles_by_memory: create_node(1000 * 1024, ...) -> Memory.from_bytes(1) passes.
-        # Let's assume create_node takes Bytes or KB consistently.
-        # If I give 100*1024*1024 bytes = 100MB.
-        # Model storage = 300KB.
-        # So capacity is definitely not an issue.
-        hidden_size=1000,
-        supports_tensor=True,
-    )
-
-    cycles = topology.get_cycles()
-    # Depending on how get_cycles works and order of addition, we might get multiple cycles.
-    # filtering by memory usually done in master.
-    # Here we just pick one.
-    selected_cycle = cycles[0]
-
-    # act
-    shard_assignments = get_shard_assignments(
-        model_meta, selected_cycle, Sharding.Pipeline
-    )
-
-    # assert
-    runner_id_a = shard_assignments.node_to_runner[node_a_id]
-    runner_id_b = shard_assignments.node_to_runner[node_b_id]
-    runner_id_c = shard_assignments.node_to_runner[node_c_id]
-
-    # Get layer counts
-    layers_a = (
-        shard_assignments.runner_to_shard[runner_id_a].end_layer
-        - shard_assignments.runner_to_shard[runner_id_a].start_layer
-    )
-    layers_b = (
-        shard_assignments.runner_to_shard[runner_id_b].end_layer
-        - shard_assignments.runner_to_shard[runner_id_b].start_layer
-    )
-    layers_c = (
-        shard_assignments.runner_to_shard[runner_id_c].end_layer
-        - shard_assignments.runner_to_shard[runner_id_c].start_layer
-    )
-
-    # Check total
-    assert layers_a + layers_b + layers_c == 30
-
-    # Check that the fastest node (A with 400GB/s) gets saturated first.
-    # With strict greedy assignment and plenty of RAM:
-    # 1. Reserve: A=1, B=1, C=1. Remaining=27.
-    # 2. Sort: [A, B, C]
-    # 3. A takes min(remaining=27, capacity=huge) = 27.
-    # 4. A=28, B=1, C=1.
-
-    assert layers_a == 28
-    assert layers_b == 1
-    assert layers_c == 1

src/exo/shared/logging.py

@@ -29,11 +29,6 @@ class _InterceptHandler(logging.Handler):
 
 def logger_setup(log_file: Path | None, verbosity: int = 0):
     """Set up logging for this process - formatting, file handles, verbosity and output"""
-
-    logging.getLogger("exo_pyo3_bindings").setLevel(logging.WARNING)
-    logging.getLogger("httpx").setLevel(logging.WARNING)
-    logging.getLogger("httpcore").setLevel(logging.WARNING)
-
     logger.remove()
 
     # replace all stdlib loggers with _InterceptHandlers that log to loguru

src/exo/shared/models/model_cards.py

@@ -1,5 +1,8 @@
-from exo.shared.types.memory import Memory
+from anyio import Path, open_file
+import tomlkit
+
 from exo.shared.types.models import ModelId, ModelMetadata
+from exo.shared.models.model_meta import get_model_meta
 from exo.utils.pydantic_ext import CamelCaseModel
 
 
@@ -11,542 +14,27 @@ class ModelCard(CamelCaseModel):
     tags: list[str]
     metadata: ModelMetadata
 
+    @staticmethod
+    async def load(path: Path) -> "ModelCard":
+        async with await open_file(path) as f:
+            data = await f.read()
+            py = tomlkit.loads(data)
+            return ModelCard.model_validate(py)
 
-MODEL_CARDS: dict[str, ModelCard] = {
-    # deepseek v3
-    "deepseek-v3.1-4bit": ModelCard(
-        short_id="deepseek-v3.1-4bit",
-        model_id=ModelId("mlx-community/DeepSeek-V3.1-4bit"),
-        name="DeepSeek V3.1 (4-bit)",
-        description="""DeepSeek V3.1 is a large language model trained on the DeepSeek V3.1 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/DeepSeek-V3.1-4bit"),
-            pretty_name="DeepSeek V3.1 (4-bit)",
-            storage_size=Memory.from_gb(378),
-            n_layers=61,
-            hidden_size=7168,
-            supports_tensor=True,
-        ),
-    ),
-    "deepseek-v3.1-8bit": ModelCard(
-        short_id="deepseek-v3.1-8bit",
-        model_id=ModelId("mlx-community/DeepSeek-V3.1-8bit"),
-        name="DeepSeek V3.1 (8-bit)",
-        description="""DeepSeek V3.1 is a large language model trained on the DeepSeek V3.1 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/DeepSeek-V3.1-8bit"),
-            pretty_name="DeepSeek V3.1 (8-bit)",
-            storage_size=Memory.from_gb(713),
-            n_layers=61,
-            hidden_size=7168,
-            supports_tensor=True,
-        ),
-    ),
-    # kimi k2
-    "kimi-k2-instruct-4bit": ModelCard(
-        short_id="kimi-k2-instruct-4bit",
-        model_id=ModelId("mlx-community/Kimi-K2-Instruct-4bit"),
-        name="Kimi K2 Instruct (4-bit)",
-        description="""Kimi K2 is a large language model trained on the Kimi K2 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Kimi-K2-Instruct-4bit"),
-            pretty_name="Kimi K2 Instruct (4-bit)",
-            storage_size=Memory.from_gb(578),
-            n_layers=61,
-            hidden_size=7168,
-            supports_tensor=True,
-        ),
-    ),
-    "kimi-k2-thinking": ModelCard(
-        short_id="kimi-k2-thinking",
-        model_id=ModelId("mlx-community/Kimi-K2-Thinking"),
-        name="Kimi K2 Thinking (4-bit)",
-        description="""Kimi K2 Thinking is the latest, most capable version of open-source thinking model.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Kimi-K2-Thinking"),
-            pretty_name="Kimi K2 Thinking (4-bit)",
-            storage_size=Memory.from_gb(658),
-            n_layers=61,
-            hidden_size=7168,
-            supports_tensor=True,
-        ),
-    ),
-    # llama-3.1
-    "llama-3.1-8b": ModelCard(
-        short_id="llama-3.1-8b",
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"),
-        name="Llama 3.1 8B (4-bit)",
-        description="""Llama 3.1 is a large language model trained on the Llama 3.1 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"),
-            pretty_name="Llama 3.1 8B (4-bit)",
-            storage_size=Memory.from_mb(4423),
-            n_layers=32,
-            hidden_size=4096,
-            supports_tensor=True,
-        ),
-    ),
-    "llama-3.1-8b-8bit": ModelCard(
-        short_id="llama-3.1-8b-8bit",
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"),
-        name="Llama 3.1 8B (8-bit)",
-        description="""Llama 3.1 is a large language model trained on the Llama 3.1 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"),
-            pretty_name="Llama 3.1 8B (8-bit)",
-            storage_size=Memory.from_mb(8540),
-            n_layers=32,
-            hidden_size=4096,
-            supports_tensor=True,
-        ),
-    ),
-    "llama-3.1-8b-bf16": ModelCard(
-        short_id="llama-3.1-8b-bf16",
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"),
-        name="Llama 3.1 8B (BF16)",
-        description="""Llama 3.1 is a large language model trained on the Llama 3.1 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"),
-            pretty_name="Llama 3.1 8B (BF16)",
-            storage_size=Memory.from_mb(16100),
-            n_layers=32,
-            hidden_size=4096,
-            supports_tensor=True,
-        ),
-    ),
-    "llama-3.1-70b": ModelCard(
-        short_id="llama-3.1-70b",
-        model_id=ModelId("mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"),
-        name="Llama 3.1 70B (4-bit)",
-        description="""Llama 3.1 is a large language model trained on the Llama 3.1 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"),
-            pretty_name="Llama 3.1 70B (4-bit)",
-            storage_size=Memory.from_mb(38769),
-            n_layers=80,
-            hidden_size=8192,
-            supports_tensor=True,
-        ),
-    ),
-    # llama-3.2
-    "llama-3.2-1b": ModelCard(
-        short_id="llama-3.2-1b",
-        model_id=ModelId("mlx-community/Llama-3.2-1B-Instruct-4bit"),
-        name="Llama 3.2 1B (4-bit)",
-        description="""Llama 3.2 is a large language model trained on the Llama 3.2 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Llama-3.2-1B-Instruct-4bit"),
-            pretty_name="Llama 3.2 1B (4-bit)",
-            storage_size=Memory.from_mb(696),
-            n_layers=16,
-            hidden_size=2048,
-            supports_tensor=True,
-        ),
-    ),
-    "llama-3.2-3b": ModelCard(
-        short_id="llama-3.2-3b",
-        model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-4bit"),
-        name="Llama 3.2 3B (4-bit)",
-        description="""Llama 3.2 is a large language model trained on the Llama 3.2 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-4bit"),
-            pretty_name="Llama 3.2 3B (4-bit)",
-            storage_size=Memory.from_mb(1777),
-            n_layers=28,
-            hidden_size=3072,
-            supports_tensor=True,
-        ),
-    ),
-    "llama-3.2-3b-8bit": ModelCard(
-        short_id="llama-3.2-3b-8bit",
-        model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-8bit"),
-        name="Llama 3.2 3B (8-bit)",
-        description="""Llama 3.2 is a large language model trained on the Llama 3.2 dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-8bit"),
-            pretty_name="Llama 3.2 3B (8-bit)",
-            storage_size=Memory.from_mb(3339),
-            n_layers=28,
-            hidden_size=3072,
-            supports_tensor=True,
-        ),
-    ),
-    # llama-3.3
-    "llama-3.3-70b": ModelCard(
-        short_id="llama-3.3-70b",
-        model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-4bit"),
-        name="Llama 3.3 70B (4-bit)",
-        description="""The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out)""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-4bit"),
-            pretty_name="Llama 3.3 70B",
-            storage_size=Memory.from_mb(38769),
-            n_layers=80,
-            hidden_size=8192,
-            supports_tensor=True,
-        ),
-    ),
-    "llama-3.3-70b-8bit": ModelCard(
-        short_id="llama-3.3-70b-8bit",
-        model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-8bit"),
-        name="Llama 3.3 70B (8-bit)",
-        description="""The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out)""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-8bit"),
-            pretty_name="Llama 3.3 70B (8-bit)",
-            storage_size=Memory.from_mb(73242),
-            n_layers=80,
-            hidden_size=8192,
-            supports_tensor=True,
-        ),
-    ),
-    "llama-3.3-70b-fp16": ModelCard(
-        short_id="llama-3.3-70b-fp16",
-        model_id=ModelId("mlx-community/llama-3.3-70b-instruct-fp16"),
-        name="Llama 3.3 70B (FP16)",
-        description="""The Meta Llama 3.3 multilingual large language model (LLM) is an instruction tuned generative model in 70B (text in/text out)""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/llama-3.3-70b-instruct-fp16"),
-            pretty_name="Llama 3.3 70B (FP16)",
-            storage_size=Memory.from_mb(137695),
-            n_layers=80,
-            hidden_size=8192,
-            supports_tensor=True,
-        ),
-    ),
-    # qwen3
-    "qwen3-0.6b": ModelCard(
-        short_id="qwen3-0.6b",
-        model_id=ModelId("mlx-community/Qwen3-0.6B-4bit"),
-        name="Qwen3 0.6B (4-bit)",
-        description="""Qwen3 0.6B is a large language model trained on the Qwen3 0.6B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-0.6B-4bit"),
-            pretty_name="Qwen3 0.6B (4-bit)",
-            storage_size=Memory.from_mb(327),
-            n_layers=28,
-            hidden_size=1024,
-            supports_tensor=False,
-        ),
-    ),
-    "qwen3-0.6b-8bit": ModelCard(
-        short_id="qwen3-0.6b-8bit",
-        model_id=ModelId("mlx-community/Qwen3-0.6B-8bit"),
-        name="Qwen3 0.6B (8-bit)",
-        description="""Qwen3 0.6B is a large language model trained on the Qwen3 0.6B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-0.6B-8bit"),
-            pretty_name="Qwen3 0.6B (8-bit)",
-            storage_size=Memory.from_mb(666),
-            n_layers=28,
-            hidden_size=1024,
-            supports_tensor=False,
-        ),
-    ),
-    "qwen3-30b": ModelCard(
-        short_id="qwen3-30b",
-        model_id=ModelId("mlx-community/Qwen3-30B-A3B-4bit"),
-        name="Qwen3 30B A3B (4-bit)",
-        description="""Qwen3 30B is a large language model trained on the Qwen3 30B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-30B-A3B-4bit"),
-            pretty_name="Qwen3 30B A3B (4-bit)",
-            storage_size=Memory.from_mb(16797),
-            n_layers=48,
-            hidden_size=2048,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-30b-8bit": ModelCard(
-        short_id="qwen3-30b-8bit",
-        model_id=ModelId("mlx-community/Qwen3-30B-A3B-8bit"),
-        name="Qwen3 30B A3B (8-bit)",
-        description="""Qwen3 30B is a large language model trained on the Qwen3 30B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-30B-A3B-8bit"),
-            pretty_name="Qwen3 30B A3B (8-bit)",
-            storage_size=Memory.from_mb(31738),
-            n_layers=48,
-            hidden_size=2048,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-80b-a3B-4bit": ModelCard(
-        short_id="qwen3-80b-a3B-4bit",
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"),
-        name="Qwen3 80B A3B (4-bit)",
-        description="""Qwen3 80B""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"),
-            pretty_name="Qwen3 80B A3B (4-bit)",
-            storage_size=Memory.from_mb(44800),
-            n_layers=48,
-            hidden_size=2048,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-80b-a3B-8bit": ModelCard(
-        short_id="qwen3-80b-a3B-8bit",
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"),
-        name="Qwen3 80B A3B (8-bit)",
-        description="""Qwen3 80B""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"),
-            pretty_name="Qwen3 80B A3B (8-bit)",
-            storage_size=Memory.from_mb(84700),
-            n_layers=48,
-            hidden_size=2048,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-80b-a3B-thinking-4bit": ModelCard(
-        short_id="qwen3-80b-a3B-thinking-4bit",
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"),
-        name="Qwen3 80B A3B Thinking (4-bit)",
-        description="""Qwen3 80B Reasoning model""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"),
-            pretty_name="Qwen3 80B A3B (4-bit)",
-            storage_size=Memory.from_mb(84700),
-            n_layers=48,
-            hidden_size=2048,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-80b-a3B-thinking-8bit": ModelCard(
-        short_id="qwen3-80b-a3B-thinking-8bit",
-        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"),
-        name="Qwen3 80B A3B Thinking (8-bit)",
-        description="""Qwen3 80B Reasoning model""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"),
-            pretty_name="Qwen3 80B A3B (8-bit)",
-            storage_size=Memory.from_mb(84700),
-            n_layers=48,
-            hidden_size=2048,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-235b-a22b-4bit": ModelCard(
-        short_id="qwen3-235b-a22b-4bit",
-        model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"),
-        name="Qwen3 235B A22B (4-bit)",
-        description="""Qwen3 235B (Active 22B) is a large language model trained on the Qwen3 235B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"),
-            pretty_name="Qwen3 235B A22B (4-bit)",
-            storage_size=Memory.from_gb(132),
-            n_layers=94,
-            hidden_size=4096,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-235b-a22b-8bit": ModelCard(
-        short_id="qwen3-235b-a22b-8bit",
-        model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"),
-        name="Qwen3 235B A22B (8-bit)",
-        description="""Qwen3 235B (Active 22B) is a large language model trained on the Qwen3 235B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"),
-            pretty_name="Qwen3 235B A22B (8-bit)",
-            storage_size=Memory.from_gb(250),
-            n_layers=94,
-            hidden_size=4096,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-coder-480b-a35b-4bit": ModelCard(
-        short_id="qwen3-coder-480b-a35b-4bit",
-        model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"),
-        name="Qwen3 Coder 480B A35B (4-bit)",
-        description="""Qwen3 Coder 480B (Active 35B) is a large language model trained on the Qwen3 Coder 480B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"),
-            pretty_name="Qwen3 Coder 480B A35B (4-bit)",
-            storage_size=Memory.from_gb(270),
-            n_layers=62,
-            hidden_size=6144,
-            supports_tensor=True,
-        ),
-    ),
-    "qwen3-coder-480b-a35b-8bit": ModelCard(
-        short_id="qwen3-coder-480b-a35b-8bit",
-        model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"),
-        name="Qwen3 Coder 480B A35B (8-bit)",
-        description="""Qwen3 Coder 480B (Active 35B) is a large language model trained on the Qwen3 Coder 480B dataset.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"),
-            pretty_name="Qwen3 Coder 480B A35B (8-bit)",
-            storage_size=Memory.from_gb(540),
-            n_layers=62,
-            hidden_size=6144,
-            supports_tensor=True,
-        ),
-    ),
-    # gpt-oss
-    "gpt-oss-120b-MXFP4-Q8": ModelCard(
-        short_id="gpt-oss-120b-MXFP4-Q8",
-        model_id=ModelId("mlx-community/gpt-oss-120b-MXFP4-Q8"),
-        name="GPT-OSS 120B (MXFP4-Q8, MLX)",
-        description="""OpenAI's GPT-OSS 120B is a 117B-parameter Mixture-of-Experts model designed for high-reasoning and general-purpose use; this variant is a 4-bit MLX conversion for Apple Silicon.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/gpt-oss-120b-MXFP4-Q8"),
-            pretty_name="GPT-OSS 120B (MXFP4-Q8, MLX)",
-            storage_size=Memory.from_kb(68_996_301),
-            n_layers=36,
-            hidden_size=2880,
-            supports_tensor=True,
-        ),
-    ),
-    "gpt-oss-20b-MXFP4-Q8": ModelCard(
-        short_id="gpt-oss-20b-MXFP4-Q8",
-        model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"),
-        name="GPT-OSS 20B (MXFP4-Q8, MLX)",
-        description="""OpenAI's GPT-OSS 20B is a medium-sized MoE model for lower-latency and local or specialized use cases; this variant is a 4-bit MLX conversion for Apple Silicon.""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"),
-            pretty_name="GPT-OSS 20B (MXFP4-Q8, MLX)",
-            storage_size=Memory.from_kb(11_744_051),
-            n_layers=24,
-            hidden_size=2880,
-            supports_tensor=True,
-        ),
-    ),
-    # glm 4.5
-    "glm-4.5-air-8bit": ModelCard(
-        # Needs to be quantized g32 or g16 to work with tensor parallel
-        short_id="glm-4.5-air-8bit",
-        model_id=ModelId("mlx-community/GLM-4.5-Air-8bit"),
-        name="GLM 4.5 Air 8bit",
-        description="""GLM 4.5 Air 8bit""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/GLM-4.5-Air-8bit"),
-            pretty_name="GLM 4.5 Air 8bit",
-            storage_size=Memory.from_gb(114),
-            n_layers=46,
-            hidden_size=4096,
-            supports_tensor=False,
-        ),
-    ),
-    "glm-4.5-air-bf16": ModelCard(
-        short_id="glm-4.5-air-bf16",
-        model_id=ModelId("mlx-community/GLM-4.5-Air-bf16"),
-        name="GLM 4.5 Air bf16",
-        description="""GLM 4.5 Air bf16""",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/GLM-4.5-Air-bf16"),
-            pretty_name="GLM 4.5 Air bf16",
-            storage_size=Memory.from_gb(214),
-            n_layers=46,
-            hidden_size=4096,
-            supports_tensor=True,
-        ),
-    ),
-    # glm 4.7
-    "glm-4.7-4bit": ModelCard(
-        short_id="glm-4.7-4bit",
-        model_id=ModelId("mlx-community/GLM-4.7-4bit"),
-        name="GLM 4.7 4bit",
-        description="GLM 4.7 4bit",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/GLM-4.7-4bit"),
-            pretty_name="GLM 4.7 4bit",
-            storage_size=Memory.from_bytes(198556925568),
-            n_layers=91,
-            hidden_size=5120,
-            supports_tensor=True,
-        ),
-    ),
-    "glm-4.7-6bit": ModelCard(
-        short_id="glm-4.7-6bit",
-        model_id=ModelId("mlx-community/GLM-4.7-6bit"),
-        name="GLM 4.7 6bit",
-        description="GLM 4.7 6bit",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/GLM-4.7-6bit"),
-            pretty_name="GLM 4.7 6bit",
-            storage_size=Memory.from_bytes(286737579648),
-            n_layers=91,
-            hidden_size=5120,
-            supports_tensor=True,
-        ),
-    ),
-    "glm-4.7-8bit-gs32": ModelCard(
-        short_id="glm-4.7-8bit-gs32",
-        model_id=ModelId("mlx-community/GLM-4.7-8bit-gs32"),
-        name="GLM 4.7 8bit (gs32)",
-        description="GLM 4.7 8bit (gs32)",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/GLM-4.7-8bit-gs32"),
-            pretty_name="GLM 4.7 8bit (gs32)",
-            storage_size=Memory.from_bytes(396963397248),
-            n_layers=91,
-            hidden_size=5120,
-            supports_tensor=True,
-        ),
-    ),
-    # minimax-m2
-    "minimax-m2.1-8bit": ModelCard(
-        short_id="minimax-m2.1-8bit",
-        model_id=ModelId("mlx-community/MiniMax-M2.1-8bit"),
-        name="MiniMax M2.1 8bit",
-        description="MiniMax M2.1 8bit",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/MiniMax-M2.1-8bit"),
-            pretty_name="MiniMax M2.1 8bit",
-            storage_size=Memory.from_bytes(242986745856),
-            n_layers=61,
-            hidden_size=3072,
-            supports_tensor=True,
-        ),
-    ),
-    "minimax-m2.1-3bit": ModelCard(
-        short_id="minimax-m2.1-3bit",
-        model_id=ModelId("mlx-community/MiniMax-M2.1-3bit"),
-        name="MiniMax M2.1 3bit",
-        description="MiniMax M2.1 3bit",
-        tags=[],
-        metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/MiniMax-M2.1-3bit"),
-            pretty_name="MiniMax M2.1 3bit",
-            storage_size=Memory.from_bytes(100086644736),
-            n_layers=61,
-            hidden_size=3072,
-            supports_tensor=True,
-        ),
-    ),
-}
+    async def save(self, path: Path):
+        async with await open_file(path, "w") as f:
+            py = self.model_dump()
+            data = tomlkit.dumps(py)  # pyright: ignore[reportUnknownMemberType]
+            await f.write(data)
+
+    @staticmethod
+    async def from_hf(model_id: str) -> "ModelCard":
+        short_name = model_id.split("/")[-1]
+        return ModelCard(
+            short_id=short_name,
+            model_id=ModelId(model_id),
+            name=short_name,
+            description=f"Custom model from {model_id}",
+            tags=[],
+            metadata=await get_model_meta(model_id),
+        )

src/exo/shared/models/model_meta.py

@@ -6,7 +6,6 @@ from huggingface_hub import model_info
 from loguru import logger
 from pydantic import BaseModel, Field
 
-from exo.shared.models.model_cards import MODEL_CARDS
 from exo.shared.types.memory import Memory
 from exo.shared.types.models import ModelId, ModelMetadata
 from exo.worker.download.download_utils import (
@@ -108,19 +107,13 @@ async def _get_model_meta(model_id: str) -> ModelMetadata:
     config_data = await get_config_data(model_id)
     num_layers = config_data.layer_count
     mem_size_bytes = await get_safetensors_size(model_id)
-    model_card = next(
-        (card for card in MODEL_CARDS.values() if card.model_id == ModelId(model_id)),
-        None,
-    )
 
     return ModelMetadata(
         model_id=ModelId(model_id),
-        pretty_name=model_card.name if model_card is not None else model_id,
+        pretty_name=model_id,
         storage_size=mem_size_bytes,
         n_layers=num_layers,
         hidden_size=config_data.hidden_size or 0,
         # TODO: all custom models currently do not support tensor. We could add a dynamic test for this?
-        supports_tensor=model_card.metadata.supports_tensor
-        if model_card is not None
-        else False,
+        supports_tensor=False,
     )

src/exo/shared/types/profiling.py

@@ -57,7 +57,6 @@ class NodePerformanceProfile(CamelCaseModel):
     chip_id: str
     friendly_name: str
     memory: MemoryPerformanceProfile
-    memory_bandwidth: int | None = None
     network_interfaces: list[NetworkInterfaceInfo] = []
     system: SystemPerformanceProfile
 

src/exo/worker/engines/mlx/utils_mlx.py

@@ -20,7 +20,6 @@ except ImportError:
 
 from mlx_lm.models.cache import KVCache, QuantizedKVCache, RotatingKVCache
 from mlx_lm.models.deepseek_v3 import DeepseekV3Model
-from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
 from exo.worker.engines.mlx.constants import (
@@ -366,8 +365,6 @@ def apply_chat_template(
         tools=chat_task_data.tools,
     )
 
-    logger.info(prompt)
-
     return prompt
 
 
@@ -399,11 +396,6 @@ def make_kv_cache(
 ) -> list[KVCache | RotatingKVCache | QuantizedKVCache]:
     assert hasattr(model, "layers")
 
-    # TODO: Do this for all models
-    if hasattr(model, "make_cache") and isinstance(model, GptOssModel):
-        logger.info("Using MLX LM's make cache")
-        return model.make_cache()  # type: ignore
-
     if max_kv_size is None:
         if KV_CACHE_BITS is None:
             logger.info("Using default KV cache")

src/exo/worker/runner/runner.py

@@ -1,15 +1,6 @@
 import time
-from collections.abc import Generator
-from functools import cache
 
 import mlx.core as mx
-from mlx_lm.models.gpt_oss import Model as GptOssModel
-from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
-    HarmonyEncodingName,
-    Role,
-    StreamableParser,
-    load_harmony_encoding,
-)
 
 from exo.shared.types.api import ChatCompletionMessageText
 from exo.shared.types.chunks import TokenChunk
@@ -162,19 +153,11 @@ def main(
                     _check_for_debug_prompts(task_params.messages[0].content)
 
                     # Generate responses using the actual MLX generation
-                    mlx_generator = mlx_generate(
+                    for response in mlx_generate(
                         model=model,
                         tokenizer=tokenizer,
                         task=task_params,
-                    )
-
-                    # GPT-OSS specific parsing to match other model formats.
-                    if isinstance(model, GptOssModel):
-                        mlx_generator = parse_gpt_oss(mlx_generator)
-
-                    # TODO: Add tool call parser here
-
-                    for response in mlx_generator:
+                    ):
                         match response:
                             case GenerationResponse():
                                 if shard_metadata.device_rank == 0:
@@ -224,43 +207,6 @@ def main(
                 break
 
 
-@cache
-def get_gpt_oss_encoding():
-    encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
-    return encoding
-
-
-def parse_gpt_oss(
-    responses: Generator[GenerationResponse],
-) -> Generator[GenerationResponse]:
-    encoding = get_gpt_oss_encoding()
-    stream = StreamableParser(encoding, role=Role.ASSISTANT)
-    thinking = False
-
-    for response in responses:
-        stream.process(response.token)
-
-        delta = stream.last_content_delta
-        ch = stream.current_channel
-
-        if ch == "analysis" and not thinking:
-            thinking = True
-            yield response.model_copy(update={"text": "<think>"})
-
-        if ch != "analysis" and thinking:
-            thinking = False
-            yield response.model_copy(update={"text": "</think>"})
-
-        if delta:
-            yield response.model_copy(update={"text": delta})
-
-        if response.finish_reason is not None:
-            if thinking:
-                yield response.model_copy(update={"text": "</think>"})
-            yield response
-            break
-
-
 EXO_RUNNER_MUST_FAIL = "EXO RUNNER MUST FAIL"
 EXO_RUNNER_MUST_OOM = "EXO RUNNER MUST OOM"
 EXO_RUNNER_MUST_TIMEOUT = "EXO RUNNER MUST TIMEOUT"

src/exo/worker/utils/net_profile.py

@@ -1,64 +1,49 @@
-import anyio
-import httpx
-from anyio import create_task_group
+import http.client
+
+from anyio import create_task_group, to_thread
 from loguru import logger
 
 from exo.shared.topology import Topology
 from exo.shared.types.common import NodeId
 
-REACHABILITY_ATTEMPTS = 3
-
 
 async def check_reachability(
     target_ip: str,
     expected_node_id: NodeId,
+    self_node_id: NodeId,
     out: dict[NodeId, set[str]],
-    client: httpx.AsyncClient,
 ) -> None:
     """Check if a node is reachable at the given IP and verify its identity."""
-    if ":" in target_ip:
-        # TODO: use real IpAddress types
-        target_ip = f"[{target_ip}]"
-    url = f"http://{target_ip}:52415/node_id"
 
-    remote_node_id = None
-    last_error = None
-
-    for _ in range(REACHABILITY_ATTEMPTS):
+    def _fetch_remote_node_id() -> NodeId | None:
+        connection = http.client.HTTPConnection(target_ip, 52415, timeout=1)
         try:
-            r = await client.get(url)
-            if r.status_code != 200:
-                await anyio.sleep(1)
-                continue
+            connection.request("GET", "/node_id")
+            response = connection.getresponse()
+            if response.status != 200:
+                return None
 
-            body = r.text.strip().strip('"')
-            if not body:
-                await anyio.sleep(1)
-                continue
+            body = response.read().decode("utf-8").strip()
 
-            remote_node_id = NodeId(body)
-            break
+            # Strip quotes if present (JSON string response)
+            if body.startswith('"') and body.endswith('"') and len(body) >= 2:
+                body = body[1:-1]
 
-        # expected failure cases
-        except (
-            httpx.TimeoutException,
-            httpx.NetworkError,
-        ):
-            await anyio.sleep(1)
-
-        # other failures should be logged on last attempt
-        except httpx.HTTPError as e:
-            last_error = e
-            await anyio.sleep(1)
-
-    if last_error is not None:
-        logger.warning(
-            f"connect error {type(last_error).__name__} from {target_ip} after {REACHABILITY_ATTEMPTS} attempts; treating as down"
-        )
+            return NodeId(body) or None
+        except OSError:
+            return None
+        except http.client.HTTPException:
+            return None
+        finally:
+            connection.close()
 
+    remote_node_id = await to_thread.run_sync(_fetch_remote_node_id)
     if remote_node_id is None:
         return
 
+    if remote_node_id == self_node_id:
+        return
+
     if remote_node_id != expected_node_id:
         logger.warning(
             f"Discovered node with unexpected node_id; "
@@ -76,33 +61,18 @@ async def check_reachable(
     topology: Topology, self_node_id: NodeId
 ) -> dict[NodeId, set[str]]:
     """Check which nodes are reachable and return their IPs."""
-
     reachable: dict[NodeId, set[str]] = {}
-
-    # these are intentionally httpx's defaults so we can tune them later
-    timeout = httpx.Timeout(timeout=5.0)
-    limits = httpx.Limits(
-        max_connections=100,
-        max_keepalive_connections=20,
-        keepalive_expiry=5,
-    )
-
-    async with (
-        httpx.AsyncClient(timeout=timeout, limits=limits) as client,
-        create_task_group() as tg,
-    ):
+    async with create_task_group() as tg:
         for node in topology.list_nodes():
             if not node.node_profile:
                 continue
-            if node.node_id == self_node_id:
-                continue
             for iface in node.node_profile.network_interfaces:
                 tg.start_soon(
                     check_reachability,
                     iface.ip_address,
                     node.node_id,
+                    self_node_id,
                     reachable,
-                    client,
                 )
 
     return reachable

src/exo/worker/utils/profile.py

@@ -4,7 +4,6 @@ import platform
 from typing import Any, Callable, Coroutine
 
 import anyio
-from anyio import to_thread
 from loguru import logger
 
 from exo.shared.types.memory import Memory
@@ -25,61 +24,8 @@ from .system_info import (
     get_friendly_name,
     get_model_and_chip,
     get_network_interfaces,
-    profile_memory_bandwidth,
 )
 
-# Module-level cache for memory bandwidth (doesn't change at runtime)
-_cached_bandwidth: int | None = None
-_bandwidth_profiled: bool = False
-_bandwidth_profiling_task: asyncio.Task[int | None] | None = None
-
-
-async def profile_bandwidth_once() -> int | None:
-    """Profile bandwidth once in a background thread and cache the result.
-
-    This function is non-blocking - it runs the profiling in a thread pool.
-    Subsequent calls return the cached result immediately.
-    """
-    global _cached_bandwidth, _bandwidth_profiled, _bandwidth_profiling_task
-
-    # Already profiled, return cached value
-    if _bandwidth_profiled:
-        return _cached_bandwidth
-
-    # Profiling already in progress, wait for it
-    if _bandwidth_profiling_task is not None:
-        return await _bandwidth_profiling_task
-
-    # Start profiling in background thread
-    async def _do_profile() -> int | None:
-        global _cached_bandwidth, _bandwidth_profiled
-        try:
-            logger.info("Starting memory bandwidth profiling in background thread...")
-            bandwidth = await to_thread.run_sync(profile_memory_bandwidth, cancellable=True)
-            _cached_bandwidth = bandwidth
-            _bandwidth_profiled = True
-            if bandwidth:
-                logger.info(f"Memory bandwidth profiled: {bandwidth / 1e9:.1f} GB/s")
-            else:
-                logger.warning("Memory bandwidth profiling returned None")
-            return bandwidth
-        except Exception as e:
-            logger.opt(exception=e).error("Memory bandwidth profiling failed")
-            _bandwidth_profiled = True  # Mark as done to avoid retrying
-            return None
-
-    _bandwidth_profiling_task = asyncio.create_task(_do_profile())
-    return await _bandwidth_profiling_task
-
-
-def get_memory_bandwidth_cached() -> int | None:
-    """Return cached bandwidth or None if not yet profiled.
-
-    This is a non-blocking synchronous function that returns immediately.
-    Call profile_bandwidth_once() first to trigger profiling.
-    """
-    return _cached_bandwidth if _bandwidth_profiled else None
-
 
 async def get_metrics_async() -> Metrics | None:
     """Return detailed Metrics on macOS or a minimal fallback elsewhere."""
@@ -125,8 +71,6 @@ async def start_polling_node_metrics(
     callback: Callable[[NodePerformanceProfile], Coroutine[Any, Any, None]],
 ):
     poll_interval_s = 1.0
-    bandwidth_profile_started = False
-
     while True:
         try:
             metrics = await get_metrics_async()
@@ -141,15 +85,6 @@ async def start_polling_node_metrics(
             # do the memory profile last to get a fresh reading to not conflict with the other memory profiling loop
             memory_profile = get_memory_profile()
 
-            # Start bandwidth profiling in background on first poll (non-blocking)
-            if not bandwidth_profile_started:
-                bandwidth_profile_started = True
-                # Fire and forget - don't await, let it run in background
-                asyncio.create_task(profile_bandwidth_once())
-
-            # Use cached bandwidth (None until profiling completes)
-            memory_bandwidth = get_memory_bandwidth_cached()
-
             await callback(
                 NodePerformanceProfile(
                     model_id=model_id,
@@ -157,7 +92,6 @@ async def start_polling_node_metrics(
                     friendly_name=friendly_name,
                     network_interfaces=network_interfaces,
                     memory=memory_profile,
-                    memory_bandwidth=memory_bandwidth,
                     system=SystemPerformanceProfile(
                         gpu_usage=metrics.gpu_usage[1],
                         temp=metrics.temp.gpu_temp_avg,

src/exo/worker/utils/system_info.py

@@ -1,6 +1,5 @@
 import socket
 import sys
-import time
 from subprocess import CalledProcessError
 
 import psutil
@@ -82,68 +81,3 @@ async def get_model_and_chip() -> tuple[str, str]:
     chip = chip_line.split(": ")[1] if chip_line else "Unknown Chip"
 
     return (model, chip)
-
-
-def profile_memory_bandwidth() -> int | None:
-    """
-    Profile device memory bandwidth using MLX GPU operations.
-
-    Uses a large array copy on the GPU to measure unified memory bandwidth.
-    Returns measured bandwidth in bytes/second, or None if MLX is unavailable.
-    """
-    try:
-        import mlx.core as mx
-
-        if not mx.metal.is_available():
-            return None
-
-        # Use 2GB buffer to better saturate memory bandwidth
-        # Use 2D shape to avoid potential issues with very large 1D arrays
-        size_bytes = 2 * 1024 * 1024 * 1024
-        side = int((size_bytes // 4) ** 0.5)  # Square 2D array of float32
-        shape = (side, side)
-        actual_bytes = side * side * 4
-        bytes_transferred = actual_bytes * 2  # read + write
-
-        # Warm-up: run the full benchmark operation multiple times to stabilize GPU
-        for _ in range(3):
-            src = mx.random.uniform(shape=shape, dtype=mx.float32)
-            mx.eval(src)
-            dst = src + 0.0
-            mx.eval(dst)
-            mx.synchronize()
-            del src, dst
-
-        # Benchmark: measure time to copy array
-        best_bandwidth = 0.0
-        num_runs = 4
-
-        for _ in range(num_runs):
-            src = mx.random.uniform(shape=shape, dtype=mx.float32)
-            mx.eval(src)
-            mx.synchronize()
-
-            # Time the copy operation (src + 0.0 forces read of src, write of dst)
-            start = time.perf_counter()
-            dst = src + 0.0
-            mx.eval(dst)
-            mx.synchronize()
-            end = time.perf_counter()
-
-            bandwidth = bytes_transferred / (end - start)
-            best_bandwidth = max(best_bandwidth, bandwidth)
-
-            del src, dst
-
-        return int(best_bandwidth)
-    except Exception:
-        return None
-
-
-def get_memory_bandwidth(_chip_id: str) -> int | None:
-    """
-    Returns measured memory bandwidth in bytes/second.
-
-    Uses MLX GPU operations for accurate unified memory bandwidth measurement.
-    """
-    return profile_memory_bandwidth()

uv.lock

@@ -236,7 +236,6 @@ dependencies = [
     { name = "exo-pyo3-bindings", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "fastapi", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "filelock", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "httpx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "huggingface-hub", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "hypercorn", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "loguru", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -248,6 +247,7 @@ dependencies = [
     { name = "pydantic", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "rustworkx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "tiktoken", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "tomlkit", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "types-aiofiles", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
 
@@ -269,7 +269,6 @@ requires-dist = [
     { name = "exo-pyo3-bindings", editable = "rust/exo_pyo3_bindings" },
     { name = "fastapi", specifier = ">=0.116.1" },
     { name = "filelock", specifier = ">=3.18.0" },
-    { name = "httpx", specifier = ">=0.28.1" },
     { name = "huggingface-hub", specifier = ">=0.33.4" },
     { name = "hypercorn", specifier = ">=0.18.0" },
     { name = "loguru", specifier = ">=0.7.3" },
@@ -281,6 +280,7 @@ requires-dist = [
     { name = "pydantic", specifier = ">=2.11.7" },
     { name = "rustworkx", specifier = ">=0.17.1" },
     { name = "tiktoken", specifier = ">=0.12.0" },
+    { name = "tomlkit", specifier = ">=0.14.0" },
     { name = "types-aiofiles", specifier = ">=24.1.0.20250708" },
 ]
 
@@ -1380,6 +1380,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/05/a1/d62dfe7376beaaf1394917e0f8e93ee5f67fea8fcf4107501db35996586b/tokenizers-0.22.2-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:38337540fbbddff8e999d59970f3c6f35a82de10053206a7562f1ea02d046fa5", size = 10033429, upload-time = "2026-01-05T10:45:14.333Z" },
 ]
 
+[[package]]
+name = "tomlkit"
+version = "0.14.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/c3/af/14b24e41977adb296d6bd1fb59402cf7d60ce364f90c890bd2ec65c43b5a/tomlkit-0.14.0.tar.gz", hash = "sha256:cf00efca415dbd57575befb1f6634c4f42d2d87dbba376128adb42c121b87064", size = 187167, upload-time = "2026-01-13T01:14:53.304Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b5/11/87d6d29fb5d237229d67973a6c9e06e048f01cf4994dee194ab0ea841814/tomlkit-0.14.0-py3-none-any.whl", hash = "sha256:592064ed85b40fa213469f81ac584f67a4f2992509a7c3ea2d632208623a3680", size = 39310, upload-time = "2026-01-13T01:14:51.965Z" },
+]
+
 [[package]]
 name = "tqdm"
 version = "4.67.1"