src/config/models.yaml

# HF Inference API Model Configuration
# Hugging Face Inference API / Inference Providers 経由で利用可能なモデルを定義
# 価格: 2026/03時点の公開価格 (USD per 1M tokens, 目安)

default_model: meta-llama/Llama-3.3-70B-Instruct
fallback_order:
  - meta-llama/Llama-3.3-70B-Instruct
  - Qwen/Qwen2.5-72B-Instruct
  - mistralai/Mistral-7B-Instruct-v0.3

models:
  # ============================================
  # Meta Llama 3.3 (テキスト/JSON向け主力)
  # ============================================
  meta-llama/Llama-3.3-70B-Instruct:
    provider: huggingface
    task: text-generation
    supports_json: true
    supports_structured_output: false
    supports_images: false
    description: "Llama 3.3 70B - 高精度テキスト生成"
    pricing:
      input: 0.30
      output: 0.60
    recommended_for:
      - default
      - high_accuracy
      - complex_reasoning
      - japanese

  meta-llama/Llama-3.1-8B-Instruct:
    provider: huggingface
    task: text-generation
    supports_json: true
    supports_structured_output: false
    supports_images: false
    description: "Llama 3.1 8B - 軽量・高速テキスト生成"
    pricing:
      input: 0.06
      output: 0.12
    recommended_for:
      - speed
      - cost_efficiency

  # ============================================
  # Qwen 2.5 (日本語対応・JSON向け)
  # ============================================
  Qwen/Qwen2.5-72B-Instruct:
    provider: huggingface
    task: text-generation
    supports_json: true
    supports_structured_output: false
    supports_images: false
    description: "Qwen 2.5 72B - 多言語対応（日本語強め）"
    pricing:
      input: 0.30
      output: 0.60
    recommended_for:
      - high_accuracy
      - japanese
      - long_context

  Qwen/Qwen2.5-7B-Instruct:
    provider: huggingface
    task: text-generation
    supports_json: true
    supports_structured_output: false
    supports_images: false
    description: "Qwen 2.5 7B - 軽量日本語対応"
    pricing:
      input: 0.06
      output: 0.12
    recommended_for:
      - speed
      - cost_efficiency
      - japanese

  # ============================================
  # Mistral (テキスト向け)
  # ============================================
  mistralai/Mistral-7B-Instruct-v0.3:
    provider: huggingface
    task: text-generation
    supports_json: true
    supports_structured_output: false
    supports_images: false
    description: "Mistral 7B - 軽量汎用テキスト生成"
    pricing:
      input: 0.06
      output: 0.12
    recommended_for:
      - speed
      - cost_efficiency

  mistralai/Mixtral-8x7B-Instruct-v0.1:
    provider: huggingface
    task: text-generation
    supports_json: true
    supports_structured_output: false
    supports_images: false
    description: "Mixtral 8x7B - 高精度MoEテキスト生成"
    pricing:
      input: 0.27
      output: 0.27
    recommended_for:
      - high_accuracy
      - coding

  # ============================================
  # Vision / Multimodal (画像入力対応)
  # ============================================
  meta-llama/Llama-3.2-11B-Vision-Instruct:
    provider: huggingface
    task: image-text-to-text
    supports_json: true
    supports_structured_output: false
    supports_images: true
    description: "Llama 3.2 11B Vision - 画像+テキスト入力"
    pricing:
      input: 0.18
      output: 0.18
    recommended_for:
      - vision
      - multimodal
      - image_analysis

  meta-llama/Llama-3.2-90B-Vision-Instruct:
    provider: huggingface
    task: image-text-to-text
    supports_json: true
    supports_structured_output: false
    supports_images: true
    description: "Llama 3.2 90B Vision - 高精度画像+テキスト入力"
    pricing:
      input: 0.90
      output: 0.90
    recommended_for:
      - vision
      - multimodal
      - high_accuracy

  Qwen/Qwen2.5-VL-72B-Instruct:
    provider: huggingface
    task: image-text-to-text
    supports_json: true
    supports_structured_output: false
    supports_images: true
    description: "Qwen2.5-VL 72B - 高精度ビジョン言語モデル（日本語対応）"
    pricing:
      input: 0.90
      output: 0.90
    recommended_for:
      - vision
      - multimodal
      - japanese
      - high_accuracy

  Qwen/Qwen2.5-VL-7B-Instruct:
    provider: huggingface
    task: image-text-to-text
    supports_json: true
    supports_structured_output: false
    supports_images: true
    description: "Qwen2.5-VL 7B - 軽量ビジョン言語モデル（日本語対応）"
    pricing:
      input: 0.18
      output: 0.18
    recommended_for:
      - vision
      - speed
      - cost_efficiency
      - japanese

  # ============================================
  # Image Generation (HF FLUX系)
  # ============================================
  black-forest-labs/FLUX.1-dev:
    provider: huggingface
    task: text-to-image
    supports_json: false
    supports_structured_output: false
    supports_images: false
    description: "FLUX.1 Dev - 高品質テキストから画像生成"
    recommended_for:
      - image_generation

  black-forest-labs/FLUX.1-schnell:
    provider: huggingface
    task: text-to-image
    supports_json: false
    supports_structured_output: false
    supports_images: false
    description: "FLUX.1 Schnell - 高速テキストから画像生成"
    recommended_for:
      - image_generation
      - speed

# ============================================
# Model Aliases
# ============================================
aliases:
  llama3: meta-llama/Llama-3.3-70B-Instruct
  llama3-small: meta-llama/Llama-3.1-8B-Instruct
  llama3-vision: meta-llama/Llama-3.2-11B-Vision-Instruct
  llama3-vision-large: meta-llama/Llama-3.2-90B-Vision-Instruct
  qwen: Qwen/Qwen2.5-72B-Instruct
  qwen-small: Qwen/Qwen2.5-7B-Instruct
  qwen-vision: Qwen/Qwen2.5-VL-72B-Instruct
  qwen-vision-small: Qwen/Qwen2.5-VL-7B-Instruct
  mistral: mistralai/Mistral-7B-Instruct-v0.3
  mixtral: mistralai/Mixtral-8x7B-Instruct-v0.1
  flux: black-forest-labs/FLUX.1-dev
  flux-fast: black-forest-labs/FLUX.1-schnell

# ============================================
# Environment Variables
# ============================================
# HF_TOKEN: Hugging Face API トークン（必須）
# HF_MODEL: デフォルトモデルの上書き（任意）