Supported Models

Overview

Aira supports three categories of models for consensus cases:

1. Built-in models — managed by Aira, ready to use immediately
2. Cloud provider models — accessed through AWS Bedrock, Azure OpenAI, or Google Vertex AI
3. Self-hosted models — your own deployments via vLLM, Ollama, or TGI

All models return structured output (decision, confidence, key factors, reasoning) regardless of provider.

Built-in Models

These models are available out of the box. Use BYOK to bring your own API keys, or let Aira manage keys for you.

OpenAI

Anthropic

Google

Cloud Provider Models

Access models through your existing cloud provider accounts. This is useful for compliance requirements (data residency), cost management (existing enterprise agreements), or accessing models not available as built-in.

AWS Bedrock

Configure Bedrock access to use Claude, Llama, and Mistral models through your AWS account.

{
  "provider": "bedrock",
  "config": {
    "aws_access_key_id": "AKIA...",
    "aws_secret_access_key": "...",
    "aws_region": "us-east-1"
  }
}

Available models through Bedrock include Claude (Opus, Sonnet, Haiku), Llama 3.3/4, and Mistral Large.

Azure OpenAI

Use GPT-5.x models through your Azure OpenAI deployment.

{
  "provider": "azure",
  "config": {
    "azure_endpoint": "https://your-resource.openai.azure.com",
    "azure_deployment": "gpt-5-4",
    "api_version": "2026-01-01-preview",
    "api_key": "..."
  }
}

Google Vertex AI

Access Claude and Gemini models through Google Cloud.

{
  "provider": "vertex",
  "config": {
    "project_id": "your-gcp-project",
    "region": "us-central1"
  }
}

Self-Hosted Models

Register self-hosted models served via vLLM, Ollama, or TGI using the Custom Models API. Self-hosted models have dedicated support with per-model system prompts and constrained decoding for guaranteed schema compliance.

Supported Self-Hosted Models

vLLM Setup

vLLM is the recommended serving engine for self-hosted models. It supports guided_json for constrained decoding, which guarantees valid structured output.

# Start vLLM with guided decoding support
vllm serve meta-llama/Llama-4-Scout-17B-16E-Instruct \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 4 \
  --enable-auto-tool-choice

Then register the model in Aira:

curl -X POST https://api.airaproof.com/api/v1/models/custom \
  -H "Authorization: Bearer aira_live_xxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Llama 4 Scout (self-hosted)",
    "model_id": "llama-4-scout",
    "endpoint_url": "https://your-gpu-server:8000/v1/chat/completions",
    "timeout_ms": 60000
  }'

Ollama Setup

For smaller deployments or development environments:

# Pull and run a model
ollama pull llama3.3:70b
ollama serve

Register with the Ollama-compatible endpoint:

curl -X POST https://api.airaproof.com/api/v1/models/custom \
  -H "Authorization: Bearer aira_live_xxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Llama 3.3 70B (Ollama)",
    "model_id": "llama-3.3-70b-ollama",
    "endpoint_url": "http://your-server:11434/v1/chat/completions",
    "timeout_ms": 60000
  }'

Structured Output by Provider

Each provider handles structured output differently. Aira abstracts this so all models return the same decision schema.

For best-effort providers, Aira validates the response and retries once if the output is malformed. If the retry also fails, the model returns a model_error result.

Model Selection Recommendations

High-Stakes Decisions (Finance, Healthcare, Legal)

Use 3-5 models from different providers for maximum independence:

{
  "models": ["gpt-5.4", "claude-opus-4-6", "gemini-3.1-pro"]
}

Consider adding o3 when the decision involves complex multi-step reasoning.

Cost-Optimized Cases

Use faster, cheaper models for high-volume, lower-risk decisions:

{
  "models": ["gpt-5-mini", "claude-haiku-4-5", "gemini-3.1-flash-lite"]
}

Maximum Coverage

Mix commercial and self-hosted models for provider diversity:

{
  "models": ["gpt-5.4", "claude-sonnet-4-6", "custom:llama-4-scout"]
}

Reasoning-Heavy Queries

Include reasoning-focused models for complex analytical decisions:

{
  "models": ["o3", "claude-opus-4-6", "gpt-5.4"]
}