Supported Models

Overview

Aira supports four categories of models:

1. Built-in models — managed by Aira, ready to use immediately
2. BYOM (Bring Your Own Model) — any OpenAI-compatible endpoint with tool calling support
3. Cloud provider models — accessed through AWS Bedrock, Azure OpenAI, or Google Vertex AI
4. Self-hosted models — your own deployments via vLLM, Ollama, or TGI

All models get the same tool-calling experience in Ask Aira chat and return structured output (decision, confidence, key factors, reasoning) in consensus cases.

Built-in Models

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

Anthropic

OpenAI

Google

DeepSeek

xAI

Mistral

Moonshot (Kimi)

Alibaba (Qwen)

BYOM — Bring Your Own Model

Register any model accessible via an OpenAI-compatible /v1/chat/completions endpoint. This includes hosted API providers, self-hosted models via vLLM/Ollama/TGI, or any custom endpoint.

Aira uses the standard OpenAI function-calling format for tool use. Models that support it get the same full tool-calling experience as built-in models — including multi-step reasoning with all Ask Aira tools.

Verified Models

These open-weight models have been verified to work with Aira's tool calling and structured output:

Any OpenAI-compatible endpoint serving these models will work — whether self-hosted or via a hosted provider.

How to Register

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",
    "model_id": "llama-3.3-70b",
    "endpoint_url": "https://your-endpoint/v1/chat/completions",
    "auth_header": "Bearer your-api-key",
    "timeout_ms": 30000
  }'

Or register from the dashboard: Models → Register Model → Custom Endpoint.

Will my custom model work?

Yes, if it meets one requirement: your endpoint accepts OpenAI-compatible POST /v1/chat/completions requests with a messages array and returns a JSON response with the content at choices[0].message.content.

This covers:

• vLLM — native OpenAI compatibility
• Ollama — OpenAI compatibility at /v1/chat/completions
• TGI — OpenAI-compatible mode
• Any hosted provider — Together AI, Fireworks, Groq, Replicate, etc.
• Your own fine-tuned model — as long as the serving layer is OpenAI-compatible

Aira sends a system prompt asking the model to return a JSON decision with decision, confidence, key_factors, and reasoning fields. Models that follow instructions well (7B+) handle this reliably. For smaller models or models that struggle with structured output, Aira validates the response and retries once.

What if my endpoint has a different format? Use the response_schema.content_path field to tell Aira where to find the response text. For example, if your endpoint returns {"result": {"text": "..."}}, set content_path to "result.text".

What if I'm fine-tuning my own model? Aira doesn't require special training. Any instruction-following model works. For best results, ensure your model can output valid JSON when asked.

Cloud Provider Models

Access models through your existing cloud provider accounts. Useful for data residency compliance, enterprise agreements, or accessing models not available as built-in.

AWS Bedrock

{
  "provider": "bedrock",
  "credentials": {
    "type": "aws",
    "access_key_id": "AKIA...",
    "secret_access_key": "...",
    "region": "us-east-1"
  }
}

Available: Claude (Opus 4.8, 4.7, 4.6, Sonnet 4.6), Llama 3.3/4, Mistral Large.

Azure OpenAI

{
  "provider": "azure",
  "credentials": {
    "type": "azure",
    "endpoint": "https://your-resource.openai.azure.com",
    "api_key": "...",
    "api_version": "2024-10-21"
  }
}

Google Vertex AI

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

Self-Hosted Models

Host models on your own infrastructure using vLLM, Ollama, or TGI. Register them as custom models — they get the same tool-calling support as any BYOM model.

Recommended for Self-Hosting

Frontier (multi-GPU)

General purpose (single A100/H100)

Reasoning

Code

Efficient (single consumer GPU)

vLLM Setup (Recommended)

vLLM supports OpenAI-compatible tool calling with constrained decoding for guaranteed valid JSON:

vllm serve meta-llama/Llama-4-Maverick-17B-128E-Instruct \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 4 \
  --enable-auto-tool-choice \
  --tool-call-parser llama4_pythonic

Then register:

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 Maverick (self-hosted)",
    "model_id": "llama-4-maverick",
    "endpoint_url": "https://your-gpu-server:8000/v1/chat/completions",
    "timeout_ms": 60000
  }'

Ollama Setup

For development and smaller deployments:

ollama pull llama3.3:70b
ollama serve

Register with the Ollama 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",
    "endpoint_url": "http://your-server:11434/v1/chat/completions",
    "timeout_ms": 60000
  }'

Capabilities Matrix

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+ models from different providers for maximum independence:

{
  "models": ["claude-opus-4-8", "gpt-5.5", "gemini-3.5-flash"]
}

PR Code Review (Consensus)

Two frontier models for high-confidence findings:

{
  "models": ["claude-sonnet-4-6", "claude-opus-4-8"]
}

Or cross-provider consensus:

{
  "models": ["claude-sonnet-4-6", "gpt-5.5", "deepseek-v4-pro"]
}

Cost-Optimized

Cheaper models for high-volume, lower-risk decisions:

{
  "models": ["claude-haiku-4-5", "gpt-5-mini", "deepseek-v4-pro"]
}

Maximum Provider Diversity

Mix commercial and open-source:

{
  "models": ["gpt-5.5", "claude-sonnet-4-6", "deepseek-v4-pro"]
}

Open-Source Only (Self-Hosted)

Full data sovereignty:

{
  "models": ["custom:gemma-4-31b", "custom:llama-4-maverick", "deepseek-v4-pro"]
}