/ mlflow / utils / model_catalog / gradient_ai.json
gradient_ai.json
  1  {
  2    "schema_version": "1.0",
  3    "models": {
  4      "alibaba-qwen3-32b": {
  5        "mode": "chat",
  6        "context_window": {
  7          "max_tokens": 2048
  8        },
  9        "capabilities": {
 10          "function_calling": false,
 11          "vision": false,
 12          "reasoning": false,
 13          "prompt_caching": false,
 14          "response_schema": false
 15        },
 16        "last_updated_at": "2026-04-24"
 17      },
 18      "anthropic-claude-3-opus": {
 19        "mode": "chat",
 20        "context_window": {
 21          "max_tokens": 1024
 22        },
 23        "pricing": {
 24          "input_per_million_tokens": 15.0,
 25          "output_per_million_tokens": 75.0
 26        },
 27        "capabilities": {
 28          "function_calling": false,
 29          "vision": false,
 30          "reasoning": false,
 31          "prompt_caching": false,
 32          "response_schema": false
 33        },
 34        "last_updated_at": "2026-04-24"
 35      },
 36      "anthropic-claude-3.5-haiku": {
 37        "mode": "chat",
 38        "context_window": {
 39          "max_tokens": 1024
 40        },
 41        "pricing": {
 42          "input_per_million_tokens": 0.8,
 43          "output_per_million_tokens": 4.0
 44        },
 45        "capabilities": {
 46          "function_calling": false,
 47          "vision": false,
 48          "reasoning": false,
 49          "prompt_caching": false,
 50          "response_schema": false
 51        },
 52        "last_updated_at": "2026-04-24"
 53      },
 54      "anthropic-claude-3.5-sonnet": {
 55        "mode": "chat",
 56        "context_window": {
 57          "max_tokens": 1024
 58        },
 59        "pricing": {
 60          "input_per_million_tokens": 3.0,
 61          "output_per_million_tokens": 15.0
 62        },
 63        "capabilities": {
 64          "function_calling": false,
 65          "vision": false,
 66          "reasoning": false,
 67          "prompt_caching": false,
 68          "response_schema": false
 69        },
 70        "last_updated_at": "2026-04-24"
 71      },
 72      "anthropic-claude-3.7-sonnet": {
 73        "mode": "chat",
 74        "context_window": {
 75          "max_tokens": 1024
 76        },
 77        "pricing": {
 78          "input_per_million_tokens": 3.0,
 79          "output_per_million_tokens": 15.0
 80        },
 81        "capabilities": {
 82          "function_calling": false,
 83          "vision": false,
 84          "reasoning": false,
 85          "prompt_caching": false,
 86          "response_schema": false
 87        },
 88        "last_updated_at": "2026-04-24"
 89      },
 90      "deepseek-r1-distill-llama-70b": {
 91        "mode": "chat",
 92        "context_window": {
 93          "max_tokens": 8000
 94        },
 95        "pricing": {
 96          "input_per_million_tokens": 0.99,
 97          "output_per_million_tokens": 0.99
 98        },
 99        "capabilities": {
100          "function_calling": false,
101          "vision": false,
102          "reasoning": false,
103          "prompt_caching": false,
104          "response_schema": false
105        },
106        "last_updated_at": "2026-04-24"
107      },
108      "llama3-8b-instruct": {
109        "mode": "chat",
110        "context_window": {
111          "max_tokens": 512
112        },
113        "pricing": {
114          "input_per_million_tokens": 0.2,
115          "output_per_million_tokens": 0.2
116        },
117        "capabilities": {
118          "function_calling": false,
119          "vision": false,
120          "reasoning": false,
121          "prompt_caching": false,
122          "response_schema": false
123        },
124        "last_updated_at": "2026-04-24"
125      },
126      "llama3.3-70b-instruct": {
127        "mode": "chat",
128        "context_window": {
129          "max_tokens": 2048
130        },
131        "pricing": {
132          "input_per_million_tokens": 0.65,
133          "output_per_million_tokens": 0.65
134        },
135        "capabilities": {
136          "function_calling": false,
137          "vision": false,
138          "reasoning": false,
139          "prompt_caching": false,
140          "response_schema": false
141        },
142        "last_updated_at": "2026-04-24"
143      },
144      "mistral-nemo-instruct-2407": {
145        "mode": "chat",
146        "context_window": {
147          "max_tokens": 512
148        },
149        "pricing": {
150          "input_per_million_tokens": 0.3,
151          "output_per_million_tokens": 0.3
152        },
153        "capabilities": {
154          "function_calling": false,
155          "vision": false,
156          "reasoning": false,
157          "prompt_caching": false,
158          "response_schema": false
159        },
160        "last_updated_at": "2026-04-24"
161      },
162      "openai-gpt-4o": {
163        "mode": "chat",
164        "context_window": {
165          "max_tokens": 16384
166        },
167        "capabilities": {
168          "function_calling": false,
169          "vision": false,
170          "reasoning": false,
171          "prompt_caching": false,
172          "response_schema": false
173        },
174        "last_updated_at": "2026-04-24"
175      },
176      "openai-gpt-4o-mini": {
177        "mode": "chat",
178        "context_window": {
179          "max_tokens": 16384
180        },
181        "capabilities": {
182          "function_calling": false,
183          "vision": false,
184          "reasoning": false,
185          "prompt_caching": false,
186          "response_schema": false
187        },
188        "last_updated_at": "2026-04-24"
189      },
190      "openai-o3": {
191        "mode": "chat",
192        "context_window": {
193          "max_tokens": 100000
194        },
195        "pricing": {
196          "input_per_million_tokens": 2.0,
197          "output_per_million_tokens": 8.0
198        },
199        "capabilities": {
200          "function_calling": false,
201          "vision": false,
202          "reasoning": false,
203          "prompt_caching": false,
204          "response_schema": false
205        },
206        "last_updated_at": "2026-04-24"
207      },
208      "openai-o3-mini": {
209        "mode": "chat",
210        "context_window": {
211          "max_tokens": 100000
212        },
213        "pricing": {
214          "input_per_million_tokens": 1.1,
215          "output_per_million_tokens": 4.4
216        },
217        "capabilities": {
218          "function_calling": false,
219          "vision": false,
220          "reasoning": false,
221          "prompt_caching": false,
222          "response_schema": false
223        },
224        "last_updated_at": "2026-04-24"
225      }
226    }
227  }