NexusRaven V2 13B Surpasses GPT-4 in Function Calling for Single, Nested and Parallel Calls

And no, it didn’t leverage any GPT-3.5/4 generated data and can generalize to tools never seen during model training!

Too good to be true? Before we get hands-on with this beast, let me explain why this is a big deal.

Join our next cohort: Full-stack GenAI SaaS Product in 4 weeks!

Translating English instructions into executable code isn’t new. But for the first time, an open-source model isn’t just matching, but surpassing a commercial giant like GPT-4. And we’re not talking simple function calls — NexusRaven v2 13B handles nested and composite functions with ease.

Derived from Meta’s CodeLlama-13B-instruct, NexusRaven V2 13B is fine-tuned with data exclusively from open-source code. It opens up endless opportunities for local and cloud based applications.

What’s also really cool is that the team behind it has set up a leaderboard on Huggingface, showcasing a diverse array of function-calling scenarios!

In this article, I’ll walk you through:

• Local setup for NexusRaven v2 13B
• Function Calling with NexusRaven v2 13B
• Benchmarks for NexusRaven v2 13B vs. GPT-4 1106 and GPT-3.5
• Use-case ideas
• Resources in case you want to dive deeper