diff options
| -rw-r--r-- | models.yaml | 159 |
1 files changed, 96 insertions, 63 deletions
diff --git a/models.yaml b/models.yaml index d76fc14..8b0d8e2 100644 --- a/models.yaml +++ b/models.yaml @@ -61,6 +61,13 @@ input_price: 3 output_price: 12 no_system_message: true + - name: o1 + max_input_tokens: 128000 + input_price: 15 + output_price: 60 + supports_vision: true + supports_function_calling: true + no_system_message: true - name: gpt-3.5-turbo max_input_tokens: 16385 max_output_tokens: 4096 @@ -159,16 +166,16 @@ max_input_tokens: 200000 max_output_tokens: 8192 require_max_tokens: true - input_price: 1 - output_price: 5 + input_price: 0.8 + output_price: 4 supports_vision: true supports_function_calling: true - name: claude-3-5-haiku-20241022 max_input_tokens: 200000 max_output_tokens: 8192 require_max_tokens: true - input_price: 1 - output_price: 5 + input_price: 0.8 + output_price: 4 supports_vision: true supports_function_calling: true - name: claude-3-opus-20240229 @@ -487,8 +494,8 @@ max_input_tokens: 200000 max_output_tokens: 8192 require_max_tokens: true - input_price: 1 - output_price: 5 + input_price: 0.8 + output_price: 4 supports_vision: true supports_function_calling: true - name: claude-3-opus@20240229 @@ -570,8 +577,8 @@ max_input_tokens: 200000 max_output_tokens: 8192 require_max_tokens: true - input_price: 1 - output_price: 5 + input_price: 0.8 + output_price: 4 supports_vision: true supports_function_calling: true - name: anthropic.claude-3-opus-20240229-v1:0 @@ -598,41 +605,48 @@ output_price: 1.25 supports_vision: true supports_function_calling: true + - name: us.meta.llama3-3-70b-instruct-v1:0 + max_input_tokens: 128000 + max_output_tokens: 8192 + require_max_tokens: true + input_price: 0.72 + output_price: 0.72 + supports_function_calling: true - name: meta.llama3-1-405b-instruct-v1:0 max_input_tokens: 128000 max_output_tokens: 4096 require_max_tokens: true - input_price: 5.32 - output_price: 16 + input_price: 2.4 + output_price: 2.4 supports_function_calling: true - name: meta.llama3-1-70b-instruct-v1:0 max_input_tokens: 128000 - max_output_tokens: 2048 + max_output_tokens: 8192 require_max_tokens: true - input_price: 0.99 - output_price: 0.99 + input_price: 0.72 + output_price: 0.72 supports_function_calling: true - name: meta.llama3-1-8b-instruct-v1:0 max_input_tokens: 128000 - max_output_tokens: 2048 + max_output_tokens: 8192 require_max_tokens: true input_price: 0.22 output_price: 0.22 supports_function_calling: true - name: us.meta.llama3-2-90b-instruct-v1:0 max_input_tokens: 128000 - max_output_tokens: 2048 + max_output_tokens: 8192 require_max_tokens: true - input_price: 2 - output_price: 2 + input_price: 0.72 + output_price: 0.72 supports_function_calling: true supports_vision: true - name: us.meta.llama3-2-11b-instruct-v1:0 max_input_tokens: 128000 - max_output_tokens: 2048 + max_output_tokens: 8192 require_max_tokens: true - input_price: 0.35 - output_price: 0.35 + input_price: 0.16 + output_price: 0.16 supports_function_calling: true supports_vision: true - name: us.amazon.nova-pro-v1:0 @@ -746,8 +760,8 @@ input_price: 4.2 output_price: 12.6 supports_function_calling: true - - name: ernie-3.5-8k-preview - max_input_tokens: 8192 + - name: ernie-3.5-128k + max_input_tokens: 128000 input_price: 0.112 output_price: 0.28 supports_function_calling: true @@ -811,6 +825,10 @@ max_input_tokens: 1000000 input_price: 0.07 output_price: 0.28 + - name: qvq-72b-preview + max_input_tokens: 16384 + max_output_tokens: 16384 + supports_vision: true - name: qwq-32b-preview max_input_tokens: 30720 max_output_tokens: 16384 @@ -862,12 +880,6 @@ input_price: 2.1 output_price: 7.0 supports_function_calling: true - - name: hunyuan-pro - max_input_tokens: 28000 - max_output_tokens: 4096 - input_price: 4.2 - output_price: 14.0 - supports_function_calling: true - name: hunyuan-large max_input_tokens: 28000 max_output_tokens: 4096 @@ -948,7 +960,7 @@ models: - name: deepseek-chat max_input_tokens: 65536 - max_output_tokens: 4096 + max_output_tokens: 8192 input_price: 0.14 output_price: 0.28 supports_function_calling: true @@ -979,16 +991,21 @@ input_price: 0.14 output_price: 0.14 supports_function_calling: true - - name: glm-4-flash + - name: glm-4-flashx max_input_tokens: 128000 - input_price: 0 - output_price: 0 + input_price: 0.014 + output_price: 0.014 supports_function_calling: true - name: glm-4v-plus max_input_tokens: 8192 input_price: 1.4 output_price: 1.4 supports_vision: true + - name: glm-4v-flash + max_input_tokens: 8192 + input_price: 0 + output_price: 0 + supports_vision: true - name: embedding-3 type: embedding max_input_tokens: 8192 @@ -1022,7 +1039,7 @@ max_input_tokens: 200000 input_price: 1.68 output_price: 1.68 - - name: yi-vision + - name: yi-vision-v2 max_input_tokens: 16384 input_price: 0.84 output_price: 0.84 @@ -1108,7 +1125,6 @@ # Links: # - https://deepinfra.com/models -# - https://deepinfra.com/pricing - platform: deepinfra models: - name: meta-llama/Llama-3.3-70B-Instruct @@ -1117,18 +1133,18 @@ output_price: 0.40 - name: meta-llama/Meta-Llama-3.1-405B-Instruct max_input_tokens: 32000 - input_price: 1.79 - output_price: 1.79 + input_price: 0.8 + output_price: 0.8 supports_function_calling: true - name: meta-llama/Meta-Llama-3.1-70B-Instruct max_input_tokens: 128000 - input_price: 0.35 + input_price: 0.23 output_price: 0.4 supports_function_calling: true - name: meta-llama/Meta-Llama-3.1-8B-Instruct max_input_tokens: 128000 - input_price: 0.055 - output_price: 0.055 + input_price: 0.03 + output_price: 0.05 supports_function_calling: true - name: meta-llama/Llama-3.2-90B-Vision-Instruct max_input_tokens: 128000 @@ -1140,33 +1156,38 @@ output_price: 0.055 - name: mistralai/Mistral-Nemo-Instruct-2407 max_input_tokens: 128000 - input_price: 0.13 - output_price: 0.13 + input_price: 0.035 + output_price: 0.08 - name: google/gemma-2-27b-it max_input_tokens: 8192 input_price: 0.27 output_price: 0.27 - name: google/gemma-2-9b-it max_input_tokens: 8192 - input_price: 0.06 + input_price: 0.03 output_price: 0.06 - name: Qwen/Qwen2.5-72B-Instruct max_input_tokens: 32768 - input_price: 0.35 + input_price: 0.23 output_price: 0.40 supports_function_calling: true - name: Qwen/Qwen2.5-Coder-32B-Instruct max_input_tokens: 32768 - input_price: 0.18 - output_price: 0.18 + input_price: 0.07 + output_price: 0.16 + - name: Qwen/QVQ-72B-Preview + max_input_tokens: 32768 + input_price: 0.25 + output_price: 0.50 + supports_vision: true - name: Qwen/QwQ-32B-Preview max_input_tokens: 32768 - input_price: 0.15 - output_price: 0.60 + input_price: 0.12 + output_price: 0.18 - name: nvidia/Llama-3.1-Nemotron-70B-Instruct max_input_tokens: 128000 - input_price: 0.35 - output_price: 0.40 + input_price: 0.12 + output_price: 0.30 supports_function_calling: true - name: BAAI/bge-large-en-v1.5 type: embedding @@ -1337,6 +1358,13 @@ input_price: 3 output_price: 12 no_system_message: true + - name: openai/o1 + max_input_tokens: 128000 + input_price: 15 + output_price: 60 + supports_vision: true + supports_function_calling: true + no_system_message: true - name: openai/gpt-3.5-turbo max_input_tokens: 16385 input_price: 0.5 @@ -1380,8 +1408,8 @@ max_input_tokens: 200000 max_output_tokens: 8192 require_max_tokens: true - input_price: 1 - output_price: 5 + input_price: 0.8 + output_price: 4 supports_vision: true supports_function_calling: true - name: anthropic/claude-3-opus @@ -1410,21 +1438,21 @@ supports_function_calling: true - name: meta-llama/llama-3.3-70b-instruct max_input_tokens: 131072 - input_price: 0.88 - output_price: 0.88 + input_price: 0.12 + output_price: 0.3 - name: meta-llama/llama-3.1-405b-instruct max_input_tokens: 131072 - input_price: 2.8 - output_price: 2.8 + input_price: 0.8 + output_price: 0.8 supports_function_calling: true - name: meta-llama/llama-3.1-70b-instruct max_input_tokens: 131072 - input_price: 0.34 - output_price: 0.39 + input_price: 0.12 + output_price: 0.3 supports_function_calling: true - name: meta-llama/llama-3.1-8b-instruct max_input_tokens: 131072 - input_price: 0.05 + input_price: 0.02 output_price: 0.05 - name: meta-llama/llama-3.2-90b-vision-instruct max_input_tokens: 131072 @@ -1451,11 +1479,6 @@ input_price: 0.1 output_price: 0.1 supports_function_calling: true - - name: mistralai/ministral-3b - max_input_tokens: 128000 - input_price: 0.04 - output_price: 0.04 - supports_function_calling: true - name: mistralai/mistral-nemo max_input_tokens: 128000 input_price: 0.13 @@ -1643,6 +1666,11 @@ input_price: 0.5782 output_price: 0.5782 supports_vision: true + - name: Qwen/QVQ-72B-Preview + max_input_tokens: 32768 + input_price: 1.386 + output_price: 1.386 + supports_vision: true - name: Qwen/QwQ-32B-Preview max_input_tokens: 32768 input_price: 0.176 @@ -1660,6 +1688,11 @@ input_price: 0.186 output_price: 0.186 supports_function_calling: true + - name: deepseek-ai/deepseek-vl2 + max_input_tokens: 32768 + input_price: 0.138 + output_price: 0.138 + supports_vision: true - name: BAAI/bge-large-en-v1.5 type: embedding input_price: 0 |