{
"models": [
{
"name": "Llama 3.1 8B",
"family": "llama",
"params_billions": 8.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192, 131072]
},
{
"name": "Llama 3.1 70B",
"family": "llama",
"params_billions": 70.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192, 131072]
},
{
"name": "Llama 3.1 405B",
"family": "llama",
"params_billions": 405.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192, 131072]
},
{
"name": "Llama 3.2 1B",
"family": "llama",
"params_billions": 1.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192, 131072]
},
{
"name": "Llama 3.2 3B",
"family": "llama",
"params_billions": 3.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192, 131072]
},
{
"name": "Mistral 7B",
"family": "mistral",
"params_billions": 7.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192, 32768]
},
{
"name": "Mixtral 8x7B",
"family": "mistral",
"params_billions": 46.7,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [32768]
},
{
"name": "Mixtral 8x22B",
"family": "mistral",
"params_billions": 141.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [65536]
},
{
"name": "Mistral Large 2 123B",
"family": "mistral",
"params_billions": 123.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [131072]
},
{
"name": "Gemma 2 2B",
"family": "gemma",
"params_billions": 2.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192]
},
{
"name": "Gemma 2 9B",
"family": "gemma",
"params_billions": 9.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192]
},
{
"name": "Gemma 2 27B",
"family": "gemma",
"params_billions": 27.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [8192]
},
{
"name": "Phi-3 Mini 3.8B",
"family": "phi",
"params_billions": 3.8,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf", "onnx"],
"context_lengths": [4096, 131072]
},
{
"name": "Phi-3 Medium 14B",
"family": "phi",
"params_billions": 14.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf", "onnx"],
"context_lengths": [4096, 131072]
},
{
"name": "Qwen 2.5 7B",
"family": "qwen",
"params_billions": 7.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [32768, 131072]
},
{
"name": "Qwen 2.5 14B",
"family": "qwen",
"params_billions": 14.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [32768, 131072]
},
{
"name": "Qwen 2.5 72B",
"family": "qwen",
"params_billions": 72.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [32768, 131072]
},
{
"name": "DeepSeek V3 671B",
"family": "deepseek",
"params_billions": 671.0,
"default_dtype": "bf16",
"formats": ["safetensors"],
"context_lengths": [131072]
},
{
"name": "DeepSeek R1 671B",
"family": "deepseek",
"params_billions": 671.0,
"default_dtype": "bf16",
"formats": ["safetensors", "gguf"],
"context_lengths": [131072]
},
{
"name": "Falcon 7B",
"family": "falcon",
"params_billions": 7.0,
"default_dtype": "bf16",
"formats": ["safetensors"],
"context_lengths": [2048]
},
{
"name": "Falcon 40B",
"family": "falcon",
"params_billions": 40.0,
"default_dtype": "bf16",
"formats": ["safetensors"],
"context_lengths": [2048]
},
{
"name": "Falcon 180B",
"family": "falcon",
"params_billions": 180.0,
"default_dtype": "bf16",
"formats": ["safetensors"],
"context_lengths": [2048]
},
{
"name": "Command R+ 104B",
"family": "cohere",
"params_billions": 104.0,
"default_dtype": "bf16",
"formats": ["safetensors"],
"context_lengths": [131072]
},
{
"name": "DBRX 132B",
"family": "databricks",
"params_billions": 132.0,
"default_dtype": "bf16",
"formats": ["safetensors"],
"context_lengths": [32768]
},
{
"name": "Stable Diffusion XL",
"family": "stability",
"params_billions": 3.5,
"default_dtype": "fp16",
"formats": ["safetensors"],
"context_lengths": []
},
{
"name": "Whisper Large V3",
"family": "whisper",
"params_billions": 1.55,
"default_dtype": "fp16",
"formats": ["safetensors", "onnx"],
"context_lengths": []
}
]
}