migrate model cards to .toml files

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/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-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-Q4"),
-            pretty_name="GPT-OSS 20B (MXFP4-Q4, 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,
     )