/ mlflow / utils / model_catalog / gmi.json
gmi.json
  1  {
  2    "schema_version": "1.0",
  3    "models": {
  4      "MiniMaxAI/MiniMax-M2.1": {
  5        "mode": "chat",
  6        "context_window": {
  7          "max_input": 196608,
  8          "max_output": 16384,
  9          "max_tokens": 16384
 10        },
 11        "pricing": {
 12          "input_per_million_tokens": 0.3,
 13          "output_per_million_tokens": 1.2
 14        },
 15        "capabilities": {
 16          "function_calling": false,
 17          "vision": false,
 18          "reasoning": false,
 19          "prompt_caching": false,
 20          "response_schema": false
 21        },
 22        "last_updated_at": "2026-04-24"
 23      },
 24      "Qwen/Qwen3-VL-235B-A22B-Instruct-FP8": {
 25        "mode": "chat",
 26        "context_window": {
 27          "max_input": 262144,
 28          "max_output": 16384,
 29          "max_tokens": 16384
 30        },
 31        "pricing": {
 32          "input_per_million_tokens": 0.3,
 33          "output_per_million_tokens": 1.4
 34        },
 35        "capabilities": {
 36          "function_calling": false,
 37          "vision": true,
 38          "reasoning": false,
 39          "prompt_caching": false,
 40          "response_schema": false
 41        },
 42        "last_updated_at": "2026-04-24"
 43      },
 44      "anthropic/claude-opus-4": {
 45        "mode": "chat",
 46        "context_window": {
 47          "max_input": 409600,
 48          "max_output": 32000,
 49          "max_tokens": 32000
 50        },
 51        "pricing": {
 52          "input_per_million_tokens": 15.0,
 53          "output_per_million_tokens": 75.0
 54        },
 55        "capabilities": {
 56          "function_calling": true,
 57          "vision": true,
 58          "reasoning": false,
 59          "prompt_caching": false,
 60          "response_schema": false
 61        },
 62        "last_updated_at": "2026-04-24"
 63      },
 64      "anthropic/claude-opus-4.5": {
 65        "mode": "chat",
 66        "context_window": {
 67          "max_input": 409600,
 68          "max_output": 32000,
 69          "max_tokens": 32000
 70        },
 71        "pricing": {
 72          "input_per_million_tokens": 5.0,
 73          "output_per_million_tokens": 25.0
 74        },
 75        "capabilities": {
 76          "function_calling": true,
 77          "vision": true,
 78          "reasoning": false,
 79          "prompt_caching": false,
 80          "response_schema": false
 81        },
 82        "last_updated_at": "2026-04-24"
 83      },
 84      "anthropic/claude-sonnet-4": {
 85        "mode": "chat",
 86        "context_window": {
 87          "max_input": 409600,
 88          "max_output": 32000,
 89          "max_tokens": 32000
 90        },
 91        "pricing": {
 92          "input_per_million_tokens": 3.0,
 93          "output_per_million_tokens": 15.0
 94        },
 95        "capabilities": {
 96          "function_calling": true,
 97          "vision": true,
 98          "reasoning": false,
 99          "prompt_caching": false,
100          "response_schema": false
101        },
102        "last_updated_at": "2026-04-24"
103      },
104      "anthropic/claude-sonnet-4.5": {
105        "mode": "chat",
106        "context_window": {
107          "max_input": 409600,
108          "max_output": 32000,
109          "max_tokens": 32000
110        },
111        "pricing": {
112          "input_per_million_tokens": 3.0,
113          "output_per_million_tokens": 15.0
114        },
115        "capabilities": {
116          "function_calling": true,
117          "vision": true,
118          "reasoning": false,
119          "prompt_caching": false,
120          "response_schema": false
121        },
122        "last_updated_at": "2026-04-24"
123      },
124      "deepseek-ai/DeepSeek-V3-0324": {
125        "mode": "chat",
126        "context_window": {
127          "max_input": 163840,
128          "max_output": 16384,
129          "max_tokens": 16384
130        },
131        "pricing": {
132          "input_per_million_tokens": 0.28,
133          "output_per_million_tokens": 0.88
134        },
135        "capabilities": {
136          "function_calling": true,
137          "vision": false,
138          "reasoning": false,
139          "prompt_caching": false,
140          "response_schema": false
141        },
142        "last_updated_at": "2026-04-24"
143      },
144      "deepseek-ai/DeepSeek-V3.2": {
145        "mode": "chat",
146        "context_window": {
147          "max_input": 163840,
148          "max_output": 16384,
149          "max_tokens": 16384
150        },
151        "pricing": {
152          "input_per_million_tokens": 0.28,
153          "output_per_million_tokens": 0.4
154        },
155        "capabilities": {
156          "function_calling": true,
157          "vision": false,
158          "reasoning": false,
159          "prompt_caching": false,
160          "response_schema": false
161        },
162        "last_updated_at": "2026-04-24"
163      },
164      "google/gemini-3-flash-preview": {
165        "mode": "chat",
166        "context_window": {
167          "max_input": 1048576,
168          "max_output": 65536,
169          "max_tokens": 65536
170        },
171        "pricing": {
172          "input_per_million_tokens": 0.5,
173          "output_per_million_tokens": 3.0
174        },
175        "capabilities": {
176          "function_calling": true,
177          "vision": true,
178          "reasoning": false,
179          "prompt_caching": false,
180          "response_schema": false
181        },
182        "last_updated_at": "2026-04-24"
183      },
184      "google/gemini-3-pro-preview": {
185        "mode": "chat",
186        "context_window": {
187          "max_input": 1048576,
188          "max_output": 65536,
189          "max_tokens": 65536
190        },
191        "pricing": {
192          "input_per_million_tokens": 2.0,
193          "output_per_million_tokens": 12.0
194        },
195        "capabilities": {
196          "function_calling": true,
197          "vision": true,
198          "reasoning": false,
199          "prompt_caching": false,
200          "response_schema": false
201        },
202        "last_updated_at": "2026-04-24"
203      },
204      "moonshotai/Kimi-K2-Thinking": {
205        "mode": "chat",
206        "context_window": {
207          "max_input": 262144,
208          "max_output": 16384,
209          "max_tokens": 16384
210        },
211        "pricing": {
212          "input_per_million_tokens": 0.8,
213          "output_per_million_tokens": 1.2
214        },
215        "capabilities": {
216          "function_calling": false,
217          "vision": false,
218          "reasoning": false,
219          "prompt_caching": false,
220          "response_schema": false
221        },
222        "last_updated_at": "2026-04-24"
223      },
224      "openai/gpt-4o": {
225        "mode": "chat",
226        "context_window": {
227          "max_input": 131072,
228          "max_output": 16384,
229          "max_tokens": 16384
230        },
231        "pricing": {
232          "input_per_million_tokens": 2.5,
233          "output_per_million_tokens": 10.0
234        },
235        "capabilities": {
236          "function_calling": true,
237          "vision": true,
238          "reasoning": false,
239          "prompt_caching": false,
240          "response_schema": false
241        },
242        "last_updated_at": "2026-04-24"
243      },
244      "openai/gpt-4o-mini": {
245        "mode": "chat",
246        "context_window": {
247          "max_input": 131072,
248          "max_output": 16384,
249          "max_tokens": 16384
250        },
251        "pricing": {
252          "input_per_million_tokens": 0.15,
253          "output_per_million_tokens": 0.6
254        },
255        "capabilities": {
256          "function_calling": true,
257          "vision": true,
258          "reasoning": false,
259          "prompt_caching": false,
260          "response_schema": false
261        },
262        "last_updated_at": "2026-04-24"
263      },
264      "openai/gpt-5": {
265        "mode": "chat",
266        "context_window": {
267          "max_input": 409600,
268          "max_output": 32000,
269          "max_tokens": 32000
270        },
271        "pricing": {
272          "input_per_million_tokens": 1.25,
273          "output_per_million_tokens": 10.0
274        },
275        "capabilities": {
276          "function_calling": true,
277          "vision": false,
278          "reasoning": false,
279          "prompt_caching": false,
280          "response_schema": false
281        },
282        "last_updated_at": "2026-04-24"
283      },
284      "openai/gpt-5.1": {
285        "mode": "chat",
286        "context_window": {
287          "max_input": 409600,
288          "max_output": 32000,
289          "max_tokens": 32000
290        },
291        "pricing": {
292          "input_per_million_tokens": 1.25,
293          "output_per_million_tokens": 10.0
294        },
295        "capabilities": {
296          "function_calling": true,
297          "vision": false,
298          "reasoning": false,
299          "prompt_caching": false,
300          "response_schema": false
301        },
302        "last_updated_at": "2026-04-24"
303      },
304      "openai/gpt-5.2": {
305        "mode": "chat",
306        "context_window": {
307          "max_input": 409600,
308          "max_output": 32000,
309          "max_tokens": 32000
310        },
311        "pricing": {
312          "input_per_million_tokens": 1.75,
313          "output_per_million_tokens": 14.0
314        },
315        "capabilities": {
316          "function_calling": true,
317          "vision": false,
318          "reasoning": false,
319          "prompt_caching": false,
320          "response_schema": false
321        },
322        "last_updated_at": "2026-04-24"
323      },
324      "zai-org/GLM-4.7-FP8": {
325        "mode": "chat",
326        "context_window": {
327          "max_input": 202752,
328          "max_output": 16384,
329          "max_tokens": 16384
330        },
331        "pricing": {
332          "input_per_million_tokens": 0.4,
333          "output_per_million_tokens": 2.0
334        },
335        "capabilities": {
336          "function_calling": false,
337          "vision": false,
338          "reasoning": false,
339          "prompt_caching": false,
340          "response_schema": false
341        },
342        "last_updated_at": "2026-04-24"
343      }
344    }
345  }