models.json

[
  {
    "name": "SmolLM2-135M-Instruct-mobile",
    "repo_id": "dispatchAI/SmolLM2-135M-Instruct-mobile",
    "chat_format": "llama-3",
    "cpu_tps": 59.7,
    "phone_tps": 46.0,
    "size_mb": 0
  },
  {
    "name": "Qwen2.5-0.5B-Instruct-mobile-int4",
    "repo_id": "dispatchAI/Qwen2.5-0.5B-Instruct-mobile-int4",
    "chat_format": "chatml",
    "cpu_tps": 12.5,
    "phone_tps": 23.2,
    "size_mb": 0
  },
  {
    "name": "Llama-3.2-1B-Instruct-Q4-mobile",
    "repo_id": "dispatchAI/Llama-3.2-1B-Instruct-Q4-mobile",
    "chat_format": "chatml",
    "cpu_tps": 11.3,
    "phone_tps": 5.4,
    "size_mb": 0
  },
  {
    "name": "Gemma-2B-Arabic-mobile",
    "repo_id": "dispatchAI/Gemma-2B-Arabic-mobile",
    "chat_format": "gemma",
    "cpu_tps": 14.8,
    "phone_tps": null,
    "size_mb": 1554.7
  },
  {
    "name": "Llama-3.2-1B-FunctionCall-mobile",
    "repo_id": "dispatchAI/Llama-3.2-1B-FunctionCall-mobile",
    "chat_format": "chatml",
    "cpu_tps": 8.9,
    "phone_tps": null,
    "size_mb": 1925.8
  },
  {
    "name": "Qwen2.5-1.5B-Instruct-mobile-int4",
    "repo_id": "dispatchAI/Qwen2.5-1.5B-Instruct-mobile-int4",
    "chat_format": "chatml",
    "cpu_tps": 7.7,
    "phone_tps": "",
    "size_mb": 1065.6
  },
  {
    "name": "TinyLlama-1.1B-Chat-mobile-int4",
    "repo_id": "dispatchAI/TinyLlama-1.1B-Chat-mobile-int4",
    "chat_format": "chatml",
    "cpu_tps": 17.6,
    "phone_tps": "",
    "size_mb": 637.8
  },
  {
    "name": "Llama-3.2-1B-Instruct-mobile",
    "repo_id": "dispatchAI/Llama-3.2-1B-Instruct-mobile",
    "chat_format": "llama-3",
    "cpu_tps": 4.9,
    "phone_tps": "",
    "size_mb": 1259.9
  },
  {
    "name": "Llama-3.2-3B-Instruct-mobile",
    "repo_id": "dispatchAI/Llama-3.2-3B-Instruct-mobile",
    "chat_format": "llama-3",
    "cpu_tps": 4.8,
    "phone_tps": "",
    "size_mb": 1925.8
  },
  {
    "name": "SmolLM2-1.7B-Instruct-mobile",
    "repo_id": "dispatchAI/SmolLM2-1.7B-Instruct-mobile",
    "chat_format": "chatml",
    "cpu_tps": 18.7,
    "phone_tps": "",
    "size_mb": 1006.7
  },
  {
    "name": "Gemma-2-2B-IT-mobile",
    "repo_id": "dispatchAI/Gemma-2-2B-IT-mobile",
    "chat_format": "gemma",
    "cpu_tps": 8.2,
    "phone_tps": "",
    "size_mb": 1629.4
  },
  {
    "name": "Phi-3.5-mini-Instruct-mobile",
    "repo_id": "dispatchAI/Phi-3.5-mini-Instruct-mobile",
    "chat_format": "chatml",
    "cpu_tps": 8.6,
    "phone_tps": "",
    "size_mb": 2282.4
  },
  {
    "name": "MiniCPM5-1B-mobile",
    "repo_id": "dispatchAI/MiniCPM5-1B-mobile",
    "chat_format": null,
    "cpu_tps": 18.1,
    "phone_tps": "",
    "size_mb": 656.2
  },
  {
    "name": "MiniCPM-V-4.6-mobile",
    "repo_id": "dispatchAI/MiniCPM-V-4.6-mobile",
    "chat_format": "chatml",
    "cpu_tps": 23.4,
    "phone_tps": "",
    "size_mb": 504.6
  },
  {
    "name": "Qwen2.5-Coder-7B-mobile",
    "repo_id": "dispatchAI/Qwen2.5-Coder-7B-mobile",
    "chat_format": "chatml",
    "cpu_tps": 3.0,
    "phone_tps": "",
    "size_mb": 4466.1
  },
  {
    "name": "Llama-3.2-3B-FunctionCall-mobile",
    "repo_id": "dispatchAI/Llama-3.2-3B-FunctionCall-mobile",
    "chat_format": "llama-3",
    "cpu_tps": 9.1,
    "phone_tps": "",
    "size_mb": 1925.8
  },
  {
    "name": "Qwen2.5-0.5B-Coder-mobile",
    "repo_id": "dispatchAI/Qwen2.5-0.5B-Coder-mobile",
    "chat_format": "chatml",
    "cpu_tps": 8.0,
    "phone_tps": "",
    "size_mb": 379.4
  },
  {
    "name": "MiniCPM5-1B-multilingual-mobile",
    "repo_id": "dispatchAI/MiniCPM5-1B-multilingual-mobile",
    "chat_format": null,
    "cpu_tps": 18.4,
    "phone_tps": "",
    "size_mb": 656.2
  },
  {
    "name": "Qwen2.5-0.5B-Chinese-mobile",
    "repo_id": "dispatchAI/Qwen2.5-0.5B-Chinese-mobile",
    "chat_format": "chatml",
    "cpu_tps": 16.4,
    "phone_tps": "",
    "size_mb": 400.6
  },
  {
    "name": "Phi-3.5-mini-instruct-Q5-mobile",
    "repo_id": "dispatchAI/Phi-3.5-mini-instruct-Q5-mobile",
    "chat_format": "chatml",
    "cpu_tps": 7.0,
    "phone_tps": "",
    "size_mb": 2684.9
  },
  {
    "name": "Qwen2.5-0.5B-Instruct-Server-mobile",
    "repo_id": "dispatchAI/Qwen2.5-0.5B-Instruct-Server-mobile",
    "chat_format": "chatml",
    "cpu_tps": 16.7,
    "phone_tps": "",
    "size_mb": 482.3
  },
  {
    "name": "TinyLlama-1.1B-Chat-Q5-mobile",
    "repo_id": "dispatchAI/TinyLlama-1.1B-Chat-Q5-mobile",
    "chat_format": "chatml",
    "cpu_tps": 10.9,
    "phone_tps": "",
    "size_mb": 746.7
  },
  {
    "name": "Gemma-2-2B-IT-Q5-mobile",
    "repo_id": "dispatchAI/Gemma-2-2B-IT-Q5-mobile",
    "chat_format": "gemma",
    "cpu_tps": 5.7,
    "phone_tps": "",
    "size_mb": 1834.2
  },
  {
    "name": "Llama-3.2-1B-Instruct-Q6-mobile",
    "repo_id": "dispatchAI/Llama-3.2-1B-Instruct-Q6-mobile",
    "chat_format": "llama-3",
    "cpu_tps": 5.0,
    "phone_tps": "",
    "size_mb": 974.5
  },
  {
    "name": "SmolLM2-360M-Instruct-mobile",
    "repo_id": "dispatchAI/SmolLM2-360M-Instruct-mobile",
    "chat_format": "chatml",
    "cpu_tps": 29.1,
    "phone_tps": "",
    "size_mb": 258.1
  },
  {
    "name": "SmolLM2-1.7B-Instruct-Q5-mobile",
    "repo_id": "dispatchAI/SmolLM2-1.7B-Instruct-Q5-mobile",
    "chat_format": "chatml",
    "cpu_tps": 10.8,
    "phone_tps": "",
    "size_mb": 1168.7
  },
  {
    "name": "Qwen2.5-Coder-1.5B-mobile",
    "repo_id": "dispatchAI/Qwen2.5-Coder-1.5B-mobile",
    "chat_format": "chatml",
    "cpu_tps": 7.8,
    "phone_tps": "",
    "size_mb": 940.4
  },
  {
    "name": "Qwen2.5-Math-1.5B-mobile",
    "repo_id": "dispatchAI/Qwen2.5-Math-1.5B-mobile",
    "chat_format": "chatml",
    "cpu_tps": 15.7,
    "phone_tps": "",
    "size_mb": 940.4
  },
  {
    "name": "Qwen2.5-1.5B-Instruct-Q5-mobile",
    "repo_id": "dispatchAI/Qwen2.5-1.5B-Instruct-Q5-mobile",
    "chat_format": "chatml",
    "cpu_tps": 4.4,
    "phone_tps": "",
    "size_mb": 1072.9
  },
  {
    "name": "Qwen2.5-1.5B-Instruct-Q8-mobile",
    "repo_id": "dispatchAI/Qwen2.5-1.5B-Instruct-Q8-mobile",
    "chat_format": "chatml",
    "cpu_tps": 8.7,
    "phone_tps": "",
    "size_mb": 1570.3
  },
  {
    "name": "Llama-3.2-3B-Instruct-Q5-mobile",
    "repo_id": "dispatchAI/Llama-3.2-3B-Instruct-Q5-mobile",
    "chat_format": "llama-3",
    "cpu_tps": 2.6,
    "phone_tps": "",
    "size_mb": 2214.6
  }
]