migrate model cards to .toml files

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

@@ -23,6 +23,7 @@ dependencies = [
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",
     "openai-harmony>=0.0.8",
+    "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/shared/models/model_cards.py

@@ -1,3 +1,6 @@
+from anyio import Path, open_file
+import tomlkit 
+
 from exo.shared.types.memory import Memory
 from exo.shared.types.models import ModelId, ModelMetadata
 from exo.utils.pydantic_ext import CamelCaseModel
@@ -11,6 +14,21 @@ 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)
+
+    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)
+
+
+
 
 MODEL_CARDS: dict[str, ModelCard] = {
     # deepseek v3
@@ -425,15 +443,15 @@ MODEL_CARDS: dict[str, ModelCard] = {
             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.""",
+    "gpt-oss-20b-4bit": ModelCard(
+        short_id="gpt-oss-20b-4bit",
+        model_id=ModelId("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=ModelMetadata(
-            model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"),
-            pretty_name="GPT-OSS 20B (MXFP4-Q8, MLX)",
+            model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q4"),
+            pretty_name="GPT-OSS 20B (MXFP4-Q4, MLX)",
             storage_size=Memory.from_kb(11_744_051),
             n_layers=24,
             hidden_size=2880,

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

@@ -1,5 +1,4 @@
 import http.client
-import time
 
 from anyio import create_task_group, to_thread
 from loguru import logger
@@ -7,8 +6,6 @@ from loguru import logger
 from exo.shared.topology import Topology
 from exo.shared.types.common import NodeId
 
-BAD_STATUSLINE_ATTEMPTS = 3
-
 
 async def check_reachability(
     target_ip: str,
@@ -18,9 +15,8 @@ async def check_reachability(
 ) -> None:
     """Check if a node is reachable at the given IP and verify its identity."""
 
-    # TODO: use an async http client
-    def _fetch_remote_node_id(*, attempt: int = 1) -> NodeId | None:
-        connection = http.client.HTTPConnection(target_ip, 52415, timeout=3)
+    def _fetch_remote_node_id() -> NodeId | None:
+        connection = http.client.HTTPConnection(target_ip, 52415, timeout=1)
         try:
             connection.request("GET", "/node_id")
             response = connection.getresponse()
@@ -36,16 +32,7 @@ async def check_reachability(
             return NodeId(body) or None
         except OSError:
             return None
-        except http.client.BadStatusLine:
-            if attempt >= BAD_STATUSLINE_ATTEMPTS:
-                logger.warning(
-                    f"BadStatusLine from {target_ip}, after {attempt} attempts, assuming connection to {expected_node_id} has dropped"
-                )
-                return None
-            time.sleep(1)
-            return _fetch_remote_node_id(attempt=attempt + 1)
-        except http.client.HTTPException as e:
-            logger.warning(f"HTTPException from {target_ip}: {type(e).__name__}: {e}")
+        except http.client.HTTPException:
             return None
         finally:
             connection.close()