Multi Model

Multi-Model Orchestration

This skill enables ARIA to dispatch tasks to multiple LLM providers beyond Claude.

Configuration Files

Multi-Model Modes

Mode A: Full Replacement (Kimi K2 replaces Claude)

All Claude calls redirect to Kimi K2. Simplest setup.

Mode B: Hybrid (Claude + Kimi K2 together) ⭐ Recommended

Use Claude as main orchestrator, Kimi K2 for subagents/background tasks. Requires claude-code-router.

Mode C: Subagent-only

Keep Claude for everything, use Kimi K2 only when explicitly requested in Task prompts.

Supported Providers

Native Claude (via Task tool)

# Built-in - no additional config needed
Task(subagent_type="general-purpose", model="opus", prompt="...")
Task(subagent_type="general-purpose", model="sonnet", prompt="...")
Task(subagent_type="general-purpose", model="haiku", prompt="...")

External Models (via MCP or Bash)

1. OpenAI (GPT-4, GPT-4o)

# Via aichat (configured in pixi.toml)
aichat -m openai:gpt-4-turbo "Your prompt here"

# Via direct API
curl https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-4-turbo", "messages": [{"role": "user", "content": "..."}]}'

2. Local Models (Ollama)

# List available models
ollama list

# Run inference
ollama run llama3.2 "Your prompt here"

# Via API
curl http://localhost:11434/api/generate \
  -d '{"model": "llama3.2", "prompt": "..."}'

3. LocalAI

# Start LocalAI server
localai run --models-path ./models

# Query (OpenAI-compatible API)
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3", "messages": [...]}'

4. vLLM (High-throughput)

# Start vLLM server
python -m vllm.entrypoints.openai.api_server \
  --model mistralai/Mixtral-8x7B-Instruct-v0.1

# Query (OpenAI-compatible)
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "mistralai/Mixtral-8x7B-Instruct-v0.1", "messages": [...]}'

5. Moonshot Kimi K2 (OpenAI-compatible)

# Using OpenAI SDK with Moonshot endpoint
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

# Kimi K2 Instruct - best for agentic tasks
response = client.chat.completions.create(
    model="kimi-k2-instruct",
    messages=[{"role": "user", "content": "Analyze this codebase..."}],
    temperature=0.6,  # Recommended for K2
)

# Kimi K2 Thinking - step-by-step reasoning with tool use
response = client.chat.completions.create(
    model="kimi-k2-thinking",
    messages=[{"role": "user", "content": "Debug this complex issue..."}],
)

# Via curl (OpenAI-compatible)
curl https://api.moonshot.ai/v1/chat/completions \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k2-instruct",
    "messages": [{"role": "user", "content": "..."}],
    "temperature": 0.6
  }'

# Via aichat (if configured)
aichat -m moonshot:kimi-k2-instruct "Your prompt here"

6. OpenRouter (Unified Gateway)

# Access Kimi K2 and 100+ models via single API
curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k2",
    "messages": [{"role": "user", "content": "..."}]
  }'

ARIA Multi-Model Dispatch Pattern

Cost-Optimized Routing

routing_strategy:
  # Use cheapest model that can handle the task
  simple_tasks:
    primary: claude.haiku      # Fast, cheap
    fallback: local.ollama     # Free, offline

  analysis_tasks:
    primary: claude.sonnet     # Balanced
    fallback: openai.gpt4      # Alternative

  complex_tasks:
    primary: claude.opus       # Best reasoning
    fallback: openai.gpt4      # Alternative

Parallel Multi-Model Execution

# Example: Get multiple perspectives on architecture decision
# Launch tasks to different models in parallel

# Claude perspective
Task(subagent_type="general-purpose", model="sonnet",
     prompt="Analyze this architecture from security standpoint...")

# OpenAI perspective (via Bash + aichat)
Bash(command='aichat -m openai:gpt-4 "Analyze this architecture..."')

# Local model (via Bash + ollama)
Bash(command='ollama run llama3.2 "Analyze this architecture..."')

Consensus Pattern

# Get agreement from multiple models before proceeding
models = ["claude.sonnet", "openai.gpt4", "local.llama3"]
responses = []

for model in models:
    response = query_model(model, prompt)
    responses.append(response)

# Require 2/3 agreement for critical decisions
consensus = check_consensus(responses, threshold=0.66)

Hybrid Mode Setup (Claude + Kimi K2)

Use Claude as the main orchestrator and Kimi K2 for subagent/background tasks.

1. Install claude-code-router

npm install -g claude-code-router

2. Create Router Config

mkdir -p ~/.claude-code-router
cp .claude/config/claude-code-router.template.json ~/.claude-code-router/config.json
# Edit with your API keys

3. Router Configuration

{
  "providers": [
    {
      "name": "anthropic",
      "api_base_url": "https://api.anthropic.com",
      "api_key": "$ANTHROPIC_API_KEY",
      "models": ["claude-opus-4-5-20251101", "claude-sonnet-4-20250514"]
    },
    {
      "name": "moonshot",
      "api_base_url": "https://api.moonshot.ai/anthropic",
      "api_key": "$MOONSHOT_API_KEY",
      "models": ["kimi-k2-thinking-turbo", "kimi-k2-instruct"]
    }
  ],
  "router": {
    "default": "anthropic,claude-sonnet-4-20250514",
    "think": "anthropic,claude-opus-4-5-20251101",
    "background": "moonshot,kimi-k2-thinking-turbo"
  }
}

4. Run with Hybrid Mode

# Activate router (sets ANTHROPIC_BASE_URL to local proxy)
ccr activate

# Run Claude Code - it will route to different models based on task
claude

5. Route Specific Subagents to Kimi K2

In Task prompts, add at the beginning:

Task(subagent_type="general-purpose",
     prompt="""<CCR-SUBAGENT-MODEL>moonshot,kimi-k2-thinking</CCR-SUBAGENT-MODEL>

     Analyze this codebase for security vulnerabilities...""")

Environment Setup

1. Copy API Key Template

# Option A: Project-local (gitignored)
cp .claude/config/env.template .claude/config/env

# Option B: User-global
cp .claude/config/env.template ~/.claude/env

# Edit with your actual API keys

2. Source Environment

# Add to ~/.bashrc or ~/.zshrc
ENV_FILE="${HOME}/.claude/env"
[ -f ".claude/config/env" ] && ENV_FILE=".claude/config/env"

if [ -f "$ENV_FILE" ]; then
  set -a
  source "$ENV_FILE"
  set +a
fi

3. Use Kimi K2 as Claude Replacement (Optional)

# Add to env file to use Kimi K2 instead of Claude:
export ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic
export ANTHROPIC_AUTH_TOKEN=${MOONSHOT_API_KEY}
export ANTHROPIC_MODEL=kimi-k2-thinking-turbo
export ANTHROPIC_DEFAULT_OPUS_MODEL=kimi-k2-thinking-turbo
export ANTHROPIC_DEFAULT_SONNET_MODEL=kimi-k2-thinking-turbo
export ANTHROPIC_DEFAULT_HAIKU_MODEL=kimi-k2-thinking-turbo
export CLAUDE_CODE_SUBAGENT_MODEL=kimi-k2-thinking-turbo

# Then run Claude Code normally - it will use Kimi K2!
claude

4. Verify Configuration

# Check Claude
echo "Claude: ${ANTHROPIC_API_KEY:0:10}..."

# Check OpenAI
echo "OpenAI: ${OPENAI_API_KEY:0:10}..."

# Check Moonshot
echo "Moonshot: ${MOONSHOT_API_KEY:0:10}..."

# Check local models
ollama list
curl -s http://localhost:8080/v1/models | jq .  # LocalAI

Model Selection Matrix

Kimi K2 Specs

• Architecture: Mixture-of-Experts (MoE)
• Total Parameters: 1 trillion
• Active Parameters: 32 billion per forward pass
• Strengths: Agentic tasks, reasoning, coding
• API: OpenAI-compatible (https://api.moonshot.ai/v1)
• Recommended temp: 0.6

Security Considerations

1. Never commit API keys - Use env files outside repo
2. Rotate keys regularly - Especially for production
3. Use local models for sensitive data - PII, proprietary code
4. Audit model access - Log which models see what data
5. Rate limit external APIs - Prevent cost overruns

Troubleshooting

Model not responding

# Check if service is running
curl -s http://localhost:11434/api/tags  # Ollama
curl -s http://localhost:8080/v1/models  # LocalAI

# Check API key
curl -s https://api.openai.com/v1/models \
  -H "Authorization: Bearer $OPENAI_API_KEY" | jq .error

Slow inference

# Use quantized models locally
ollama pull llama3.2:3b-instruct-q4_0  # Smaller, faster

# Or use vLLM for batching
python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Llama-3-8b-instruct \
  --tensor-parallel-size 2  # Multi-GPU

Related Skills

• AI Assistants - aichat, aider configuration
• Inference - LocalAI, vLLM setup
• Observability - Model metrics monitoring