/ mlflow / utils / model_catalog / lemonade.json
lemonade.json
  1  {
  2    "schema_version": "1.0",
  3    "models": {
  4      "Gemma-3-4b-it-GGUF": {
  5        "mode": "chat",
  6        "context_window": {
  7          "max_input": 128000,
  8          "max_output": 8192,
  9          "max_tokens": 8192
 10        },
 11        "pricing": {
 12          "input_per_million_tokens": 0,
 13          "output_per_million_tokens": 0
 14        },
 15        "capabilities": {
 16          "function_calling": true,
 17          "vision": false,
 18          "reasoning": false,
 19          "prompt_caching": false,
 20          "response_schema": true
 21        },
 22        "last_updated_at": "2026-04-24"
 23      },
 24      "Qwen3-4B-Instruct-2507-GGUF": {
 25        "mode": "chat",
 26        "context_window": {
 27          "max_input": 262144,
 28          "max_output": 32768,
 29          "max_tokens": 32768
 30        },
 31        "pricing": {
 32          "input_per_million_tokens": 0,
 33          "output_per_million_tokens": 0
 34        },
 35        "capabilities": {
 36          "function_calling": true,
 37          "vision": false,
 38          "reasoning": false,
 39          "prompt_caching": false,
 40          "response_schema": true
 41        },
 42        "last_updated_at": "2026-04-24"
 43      },
 44      "Qwen3-Coder-30B-A3B-Instruct-GGUF": {
 45        "mode": "chat",
 46        "context_window": {
 47          "max_input": 262144,
 48          "max_output": 32768,
 49          "max_tokens": 32768
 50        },
 51        "pricing": {
 52          "input_per_million_tokens": 0,
 53          "output_per_million_tokens": 0
 54        },
 55        "capabilities": {
 56          "function_calling": true,
 57          "vision": false,
 58          "reasoning": false,
 59          "prompt_caching": false,
 60          "response_schema": true
 61        },
 62        "last_updated_at": "2026-04-24"
 63      },
 64      "gpt-oss-120b-mxfp-GGUF": {
 65        "mode": "chat",
 66        "context_window": {
 67          "max_input": 131072,
 68          "max_output": 32768,
 69          "max_tokens": 32768
 70        },
 71        "pricing": {
 72          "input_per_million_tokens": 0,
 73          "output_per_million_tokens": 0
 74        },
 75        "capabilities": {
 76          "function_calling": true,
 77          "vision": false,
 78          "reasoning": false,
 79          "prompt_caching": false,
 80          "response_schema": true
 81        },
 82        "last_updated_at": "2026-04-24"
 83      },
 84      "gpt-oss-20b-mxfp4-GGUF": {
 85        "mode": "chat",
 86        "context_window": {
 87          "max_input": 131072,
 88          "max_output": 32768,
 89          "max_tokens": 32768
 90        },
 91        "pricing": {
 92          "input_per_million_tokens": 0,
 93          "output_per_million_tokens": 0
 94        },
 95        "capabilities": {
 96          "function_calling": true,
 97          "vision": false,
 98          "reasoning": false,
 99          "prompt_caching": false,
100          "response_schema": true
101        },
102        "last_updated_at": "2026-04-24"
103      }
104    }
105  }