model_list.yaml.example

# Model inventory — single source of truth for LiteLLM routes + llmao UX.
#
#   cp model_list.yaml.example model_list.yaml
#
# litellm.yaml includes this file. Puppet/eyaml substitutes API keys only;
# api_base is not secret (cleartext here; do not show raw bases in the UX).
# Production: do not use STORE_MODEL_IN_DB for inventory — this YAML is SoT.
# Ops: restart LiteLLM after VCS/config changes (p6 systemd notify later).

model_list:
  - model_name: gemma4-26b
    litellm_params:
      model: openai/gemma4-26b
      api_base: http://127.0.0.1:8001
      api_key: CHANGE_ME_SELFHOST_API_KEY
    model_info:
      display_name: "Gemma 4 26B-A4B (self-hosted, multimodal)"
      provider: self-host
      context_window: 131072
      license: Apache-2.0
      openness: open-weight
      weights_distribution: "google/gemma-4-26B-A4B-it (HF, BF16) · vLLM"
      training_data_provenance: "undisclosed (Google DeepMind)"
      provenance_record: absent
      self_hosted: true
      modality: text+vision
      supports_thinking: true
      thinks_by_default: false
      notes: >-
        MoE, ~4B active. General/multimodal/agentic default. Served off-host
        at BF16 on an 80GB card.

  - model_name: qwen3-8b
    litellm_params:
      model: openai/qwen3-8b
      api_base: http://127.0.0.1:8003
      api_key: CHANGE_ME_SELFHOST_API_KEY
    model_info:
      display_name: "Qwen3 8B (self-hosted, fast)"
      provider: self-host
      context_window: 40960
      license: Apache-2.0
      openness: open-weight
      weights_distribution: "Qwen/Qwen3-8B-FP8 (HF) · vLLM"
      training_data_provenance: "undisclosed (Alibaba)"
      provenance_record: absent
      self_hosted: true
      modality: text
      supports_thinking: true
      thinks_by_default: true
      notes: >-
        Fast lightweight tier for routine/cheap calls. Reasoning model
        (vLLM --reasoning-parser qwen3).