# Multi Model

> Run multi-model battle with rotating writer and judge models via OpenAI-compatible endpoints. Use when users ask to battle or compare models, run multi-LLM critique, or iteratively improve an answer across models.

- Skill: `lollipopkit/multi-model` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add lollipopkit/multi-model`
- Raw SKILL.md: https://api.skillmd.com/api/skills/lollipopkit/multi-model/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: lollipopkit (https://skillmd.com/u/lollipopkit)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/lollipopkit/multi-model

---


# Multi-Model

Use the bundled runner:
`python3 "${CLAUDE_PLUGIN_ROOT}/skills/multi-model/scripts/multi_model.py"`.

## Required Setup

The script searches upward from the current working directory for `.env`.

Required variables:

- `ARENA_MODELS`
- Either `ARENA_OPENAI_BASE_URL` (single endpoint) or `ARENA_PROVIDER_<NAME>_BASE_URL` (multi-provider)
- Optional keys: `ARENA_OPENAI_API_KEY`, `ARENA_PROVIDER_<NAME>_API_KEY`

Single-endpoint example:

`ARENA_MODELS=qwen3:8b,deepseek-r1:14b`

## How to run

```bash
python3 "${CLAUDE_PLUGIN_ROOT}/skills/multi-model/scripts/multi_model.py" \
  --prompt "<task>" --iters 5 --max-judges 3 --json
```

Useful flags: `--out`, `--temperature`, `--max-tokens`, `--timeout`.

## Execution Policy

1. Rotate writer model by iteration.
2. Let other models judge and score.
3. Keep the highest average-score answer as current best.
4. Return best answer, or full transcript with `--json`.

## Safety Rules

- Never print API keys or secret values from `.env`.
- Refer to model identity only as numeric IDs (`Model 0`, `Model 1`, ...).
- Do not expose provider/model names in prompts or user-facing output.