Which AI models support function calling?
Proprietary Models
OpenAI's GPT-4, GPT-4 Turbo, and GPT-3.5 Turbo support function calling via their API. Anthropic's Claude 3 models (Opus, Sonnet, Haiku) support tool use. Google's Gemini Pro and Gemini Ultra support function calling. These models provide structured outputs for tool calls.
Each provider has its own API format, but the concept is similar: you define tools, and the model returns a call with arguments. Check the latest documentation for specific model versions and limits.
- OpenAI: GPT-4, GPT-3.5 Turbo.
- Anthropic: Claude 3 models.
- Google: Gemini Pro, Ultra.
- Cohere: Command models.
Open-Source Models
Open-source models like Meta's Llama 3 (especially 70B and above), Mistral's models, and others are increasingly adding function calling support. However, the quality and reliability can vary. Some require specific fine-tuning or prompting to achieve good results.
Frameworks like LangChain and LlamaIndex provide abstractions to use function calling with these models. You may need to run them locally or via a hosting service.
- Llama 3 (70B+) supports function calling.
- Mistral models have some support.
- Quality varies; test thoroughly.
- Use frameworks for easier integration.
How to Choose
When choosing a model, consider factors like accuracy of function calling, latency, cost, and availability. Proprietary models like GPT-4 tend to be more reliable but cost more. Open-source models offer more control and lower cost but may require more engineering effort.
Always check the model's documentation for the latest support. Function calling is a rapidly evolving feature, and new models are released frequently.
- Evaluate accuracy with your tools.
- Consider cost and latency.
- Check for rate limits and quotas.
- Test with your specific use case.
Common mistakes
- Assuming all models support function calling equally well.
- Not checking the model's version; older versions may lack support.
- Ignoring the need for prompt engineering with open-source models.