diff options
Diffstat (limited to 'models.yaml')
| -rw-r--r-- | models.yaml | 92 |
1 files changed, 64 insertions, 28 deletions
diff --git a/models.yaml b/models.yaml index 38fb642..0799308 100644 --- a/models.yaml +++ b/models.yaml @@ -8,44 +8,45 @@ models: - name: gpt-3.5-turbo max_input_tokens: 16385 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 0.5 output_price: 1.5 - name: gpt-3.5-turbo-1106 max_input_tokens: 16385 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 1 output_price: 2 - name: gpt-4-turbo max_input_tokens: 128000 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 10 output_price: 30 supports_vision: true - name: gpt-4-turbo-preview max_input_tokens: 128000 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 10 output_price: 30 - name: gpt-4-1106-preview max_input_tokens: 128000 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 10 output_price: 30 - name: gpt-4-vision-preview max_input_tokens: 128000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 10 output_price: 30 supports_vision: true - name: gpt-4 max_input_tokens: 8192 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 30 output_price: 60 - name: gpt-4-32k max_input_tokens: 32768 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 60 output_price: 120 @@ -59,18 +60,18 @@ models: - name: gemini-1.0-pro-latest max_input_tokens: 30720 - max_output_tokens?: 2048 + max_output_tokens: 2048 input_price: 0.5 output_price: 1.5 - name: gemini-1.0-pro-vision-latest max_input_tokens: 12288 - max_output_tokens?: 4096 + max_output_tokens: 4096 input_price: 0.5 output_price: 1.5 supports_vision: true - name: gemini-1.5-pro-latest max_input_tokens: 1048576 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 7 output_price: 21 supports_vision: true @@ -85,18 +86,21 @@ - name: claude-3-opus-20240229 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 15 output_price: 75 supports_vision: true - name: claude-3-sonnet-20240229 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 3 output_price: 15 supports_vision: true - name: claude-3-haiku-20240307 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.25 output_price: 1.25 supports_vision: true @@ -140,12 +144,12 @@ models: - name: command-r max_input_tokens: 128000 - max_output_tokens?: 4000 + max_output_tokens: 4000 input_price: 0.5 output_price: 1.5 - name: command-r-plus max_input_tokens: 128000 - max_output_tokens?: 4000 + max_output_tokens: 4000 input_price: 3 output_price: 15 @@ -159,28 +163,28 @@ models: - name: llama-3-sonar-small-32k-chat max_input_tokens: 32768 - max_output_tokens?: 32768 + max_output_tokens: 32768 input_price: 0.2 output_price: 0.2 - name: llama-3-sonar-large-32k-chat max_input_tokens: 32768 - max_output_tokens?: 32768 + max_output_tokens: 32768 input_price: 0.6 output_price: 0.6 - name: llama-3-8b-instruct max_input_tokens: 8192 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 0.2 output_price: 0.2 - name: llama-3-70b-instruct max_input_tokens: 8192 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 1 output_price: 1 - name: mixtral-8x7b-instruct max_input_tokens: 16384 - max_output_tokens?: 16384 + max_output_tokens: 16384 input_price: 0.6 output_price: 0.6 @@ -195,22 +199,22 @@ models: - name: llama3-8b-8192 max_input_tokens: 8192 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 0.05 output_price: 0.10 - name: llama3-70b-8192 max_input_tokens: 8192 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 0.59 output_price: 0.79 - name: mixtral-8x7b-32768 max_input_tokens: 32768 - max_output_tokens?: 32768 + max_output_tokens: 32768 input_price: 0.27 output_price: 0.27 - name: gemma-7b-it max_input_tokens: 8192 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 0.10 output_price: 0.10 @@ -224,18 +228,18 @@ models: - name: gemini-1.0-pro max_input_tokens: 24568 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 0.125 output_price: 0.375 - name: gemini-1.0-pro-vision max_input_tokens: 14336 - max_output_tokens?: 2048 + max_output_tokens: 2048 input_price: 0.125 output_price: 0.375 supports_vision: true - name: gemini-1.5-pro-preview-0409 max_input_tokens: 1000000 - max_output_tokens?: 8192 + max_output_tokens: 8192 input_price: 2.5 output_price: 7.5 supports_vision: true @@ -250,18 +254,21 @@ - name: claude-3-opus@20240229 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 15 output_price: 75 supports_vision: true - name: claude-3-sonnet@20240229 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 3 output_price: 15 supports_vision: true - name: claude-3-haiku@20240307 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.25 output_price: 1.25 supports_vision: true @@ -277,44 +284,52 @@ - name: anthropic.claude-3-opus-20240229-v1:0 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 15 output_price: 75 supports_vision: true - name: anthropic.claude-3-sonnet-20240229-v1:0 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 3 output_price: 15 supports_vision: true - name: anthropic.claude-3-haiku-20240307-v1:0 max_input_tokens: 200000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.25 output_price: 1.25 supports_vision: true - name: meta.llama3-8b-instruct-v1:0 max_input_tokens: 8192 max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.4 output_price: 0.6 - name: meta.llama3-70b-instruct-v1:0 max_input_tokens: 8192 max_output_tokens: 4096 + pass_max_tokens: true input_price: 2.65 output_price: 3.5 - name: mistral.mistral-7b-instruct-v0:2 max_input_tokens: 32000 max_output_tokens: 8192 + pass_max_tokens: true input_price: 0.15 output_price: 0.2 - name: mistral.mixtral-8x7b-instruct-v0:1 max_input_tokens: 32000 max_output_tokens: 8192 + pass_max_tokens: true input_price: 0.45 output_price: 0.7 - name: mistral.mistral-large-2402-v1:0 max_input_tokens: 32000 max_output_tokens: 8192 + pass_max_tokens: true input_price: 8 output_price: 2.4 @@ -328,21 +343,27 @@ - name: '@cf/meta/llama-3-8b-instruct' max_input_tokens: 4096 max_output_tokens: 4096 + pass_max_tokens: true - name: '@cf/mistral/mistral-7b-instruct-v0.2-lora' max_input_tokens: 4096 max_output_tokens: 4096 + pass_max_tokens: true - name: '@cf/google/gemma-7b-it-lora' max_input_tokens: 4096 max_output_tokens: 4096 + pass_max_tokens: true - name: '@cf/qwen/qwen1.5-14b-chat-awq' max_input_tokens: 4096 max_output_tokens: 4096 + pass_max_tokens: true - name: '@hf/thebloke/deepseek-coder-6.7b-instruct-awq' max_input_tokens: 4096 max_output_tokens: 4096 + pass_max_tokens: true - name: '@hf/nexusflow/starling-lm-7b-beta' max_input_tokens: 4096 max_output_tokens: 4096 + pass_max_tokens: true - platform: replicate # docs: @@ -354,21 +375,25 @@ - name: meta/meta-llama-3-70b-instruct max_input_tokens: 8192 max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.65 output_price: 2.75 - name: meta/meta-llama-3-8b-instruct max_input_tokens: 8192 max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.05 output_price: 0.25 - name: mistralai/mistral-7b-instruct-v0.2 max_input_tokens: 32000 max_output_tokens: 8192 + pass_max_tokens: true input_price: 0.05 output_price: 0.25 - name: mistralai/mixtral-8x7b-instruct-v0.1 max_input_tokens: 32000 max_output_tokens: 8192 + pass_max_tokens: true input_price: 0.3 output_price: 1 @@ -382,26 +407,31 @@ - name: ernie-4.0-8k-preview max_input_tokens: 5120 max_output_tokens: 2048 + pass_max_tokens: true input_price: 16.8 output_price: 16.8 - name: ernie-3.5-8k-preview max_input_tokens: 5120 max_output_tokens: 2048 + pass_max_tokens: true input_price: 1.68 output_price: 1.68 - name: ernie-speed-128k max_input_tokens: 124000 max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.56 output_price: 1.12 - name: ernie-lite-8k max_input_tokens: 7168 max_output_tokens: 2048 + pass_max_tokens: true input_price: 0.42 output_price: 0.84 - name: ernie-tiny-8k max_input_tokens: 7168 max_output_tokens: 2048 + pass_max_tokens: true input_price: 0.14 output_price: 0.14 @@ -414,22 +444,22 @@ models: - name: qwen-turbo max_input_tokens: 6000 - max_output_tokens?: 1500 + max_output_tokens: 1500 input_price: 1.12 output_price: 1.12 - name: qwen-plus max_input_tokens: 30000 - max_output_tokens?: 2000 + max_output_tokens: 2000 input_price: 2.8 output_price: 2.8 - name: qwen-max max_input_tokens: 6000 - max_output_tokens?: 2000 + max_output_tokens: 2000 input_price: 16.8 output_price: 16.8 - name: qwen-max-longcontext max_input_tokens: 28000 - max_output_tokens?: 2000 + max_output_tokens: 2000 - name: qwen-vl-plus input_price: 1.12 output_price: 1.12 @@ -686,16 +716,22 @@ supports_vision: true - name: anthropic/claude-3-opus max_input_tokens: 200000 + max_output_tokens: 4096 + pass_max_tokens: true input_price: 15 output_price: 75 supports_vision: true - name: anthropic/claude-3-sonnet max_input_tokens: 200000 + max_output_tokens: 4096 + pass_max_tokens: true input_price: 3 output_price: 15 supports_vision: true - name: anthropic/claude-3-haiku max_input_tokens: 200000 + max_output_tokens: 4096 + pass_max_tokens: true input_price: 0.25 output_price: 1.25 supports_vision: true |
