Skip to content Skip to footer

Fireworks AI has unveiled Firefunction-v2: A freely accessible weights function calling model featuring function calling capacity that matches GPT4o. Interestingly, it operates at a speed that’s 2.5 times faster and costs just a tenth of the price.

Fireworks AI recently launched Firefunction-v2, an open-source function-calling model aiming to deliver superior performance in real-world applications. The model integrates with multi-turn conversations, instruction following, and parallel function calling, providing a powerful and effective solution comparable to more advanced models such as GPT-4o, but with increased speed, better functionality, and lower costs. Firefunction-v2’s robustness and efficiency make it a formidable competitor in the realm of function-calling models.

The evolution of Large Language Models (LLMs) has significantly enhanced the importance of function calling—enabling these models to interact with external APIs and therefore expanding their usefulness beyond merely handling static data. Leveraging such developments, Firefunction-v2 delivers a well-rounded solution for real-world scenarios involving multi-turn conversations, instruction following, and parallel function calling.

Compared to its peers, Firefunction-v2 has shown superior performance in function-calling tasks. Priced at $0.9 per output token—a significant reduction from GPT-4o’s hefty $15 per token—Firefunction-v2 processes at a rate of 180 tokens per second, compared to GPT-4o’s 79 tokens per second.

The development of Firefunction-v2 focused on user feedback and the necessity for a model with stellar function-calling and general task proficiency. Unlike other open-source function calling models that sacrifice general reasoning abilities for specialized performance, this model strikes a balance. Originating from the Llama3-70b-instruct base model, Firefunction-v2 was meticulously fine-tuned using a handpicked dataset containing function calling and general conversation data.

Upon evaluation, Firefunction-v2 outscored its predecessor and other models like the Llama3-70b-instruct and GPT-4o across various function-calling tasks. This reinforced its versatility and agility in tackling complex tasks.

Firefunction-v2 can reliably support up to 30 function specifications—a significant increase from Firefunction-v1’s limited capacity of only five functions. This feature proves vitally important to real-world applications, allowing the model to handle multiple API calls seamlessly and providing an uninterrupted user experience.

Firefunction-v2 is available via Fireworks AI’s platform, featuring a speed-optimized setup with an OpenAI-compatible API. This ensures a smooth integration of Firefunction-v2 into existing systems with minimal necessary modifications. A demo application and UI playground are also available for users to explore the model’s various functions and configurations.

In closing, Firefunction-v2 sets a pioneering standard for real-world AI applications by harmonizing speed, cost, and performance, a testament to Fireworks AI’s commitment to pushing the envelope within the Large Language Models capabilities sphere. By continually iterating its models based on user feedback, Fireworks AI is dedicated to providing practical, industry-leading solutions for developers. As underscored by positive reviews from the developer community and high benchmark scores, Firefunction-v2 shows significant promise in revolutionizing how function calls are integrated into AI systems.

Leave a comment

0.0/5