# Prices vary for respective API providers. Default prices may not reflect current prices.

# [Structure template]
# {PROVIDER}:
#   {MODEL_NAME}:
#     model_family: {FAMILY_NAME}
#     model_alias: {MODEL_API_NAME}
#     model_context_length: {MODEL_WINDOW}
#     model_params:
#       {custom parameter arguments}. For example:
#       max_completion_tokens: 
#       temperature:
#       top_p:
#       context_length:
#       stop:
#       extra_body:

# OLLAMA CLOUD (OpenAI-compatible API)
ollama-cloud:
  base_url: https://ollama.com/v1
  deepseek-v4-pro:cloud:
    model_family: deepseek-v4
    model_alias: deepseek-v4-pro:cloud
    model_context_length: 1000000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  deepseek-v4-flash:cloud:
    model_family: deepseek-v4
    model_alias: deepseek-v4-flash:cloud
    model_context_length: 1000000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  deepseek-v3.2:cloud:
    model_family: deepseek-v3
    model_alias: deepseek-v3.2:cloud
    model_context_length: 160000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  deepseek-v3.1:671b-cloud:
    model_family: deepseek-v3
    model_alias: deepseek-v3.1:671b-cloud
    model_context_length: 160000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  kimi-k2-thinking:cloud:
    model_family: kimi-k2
    model_alias: kimi-k2-thinking:cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  kimi-k2.7-code:cloud:
    model_family: kimi-k2
    model_alias: kimi-k2.7-code:cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  kimi-k2.6:cloud:
    model_family: kimi-k2
    model_alias: kimi-k2.6:cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  kimi-k2.5:cloud:
    model_family: kimi-k2
    model_alias: kimi-k2.5:cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  kimi-k2:1t-cloud:
    model_family: kimi-k2
    model_alias: kimi-k2:1t-cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  gemini-3-flash-preview:cloud:
    model_family: gemini-3
    model_alias: gemini-3-flash-preview:cloud
    model_context_length: 1000000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  gemma4:31b-cloud:
    model_family: gemma4
    model_alias: gemma4:31b-cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  glm-5.2:cloud:
    model_family: glm-5.2
    model_alias: glm-5.2:cloud
    model_context_length: 976000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  glm-5.1:cloud:
    model_family: glm-5.1
    model_alias: glm-5.1:cloud
    model_context_length: 198000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  glm-5:cloud:
    model_family: glm-5
    model_alias: glm-5:cloud
    model_context_length: 198000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  gpt-oss:120b-cloud:
    model_family: gpt-oss
    model_alias: gpt-oss:120b-cloud
    model_context_length: 128000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  gpt-oss:20b-cloud:
    model_family: gpt-oss
    model_alias: gpt-oss:20b-cloud
    model_context_length: 128000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  mistral-large-3:675b-cloud:
    model_family: mistral-large-3
    model_alias: mistral-large-3:675b-cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  ministral-3:14b-cloud:
    model_family: ministral-3
    model_alias: ministral-3:14b-cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  ministral-3:8b-cloud:
    model_family: ministral-3
    model_alias: ministral-3:8b-cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  ministral-3:3b-cloud:
    model_family: ministral-3
    model_alias: ministral-3:3b-cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  qwen3.5:397b-cloud:
    model_family: qwen3.5
    model_alias: qwen3.5:397b-cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  qwen3.5:cloud:
    model_family: qwen3.5
    model_alias: qwen3.5:cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  minimax-m3:cloud:
    model_family: minimax-m3
    model_alias: minimax-m3:cloud
    model_context_length: 512000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00
  nemotron-3-ultra:cloud:
    model_family: nemotron-3-ultra
    model_alias: nemotron-3-ultra:cloud
    model_context_length: 256000
    model_params:
      max_completion_tokens: 8192
      temperature: 0.0
      top_p: 1.0
    cost_usd_mtok:
      input: 0.00
      output: 0.00