Nexusflow has announced the release of a 13-billion parameter open-source generative AI model that delivers function calling capability. NexusRaven-V2 can understand human instructions and translate the instructions into precise function/API calls to use a variety of software tools.

“NexusRaven-V2 outperforms OpenAI’s GPT-4 model head-to-head for using software tools,” said Jiantao Jiao, CEO and co-founder of Nexusflow. “This validates Nexusflow’s technical capability to deliver enterprise solutions using open-source models.”

According to the company, the function calling capability lies at the core of the OpenAI Assistants API, and serves as the key to enabling copilots and agents to use software tools. Instruction-tuned from the CodeLlama-13B model, NexusRaven-V2 achieves up to a 7% higher tool use success rate than the latest OpenAI GPT-4 model on human-curated general software tool use benchmarks.

Nexusflow’s data curation pipeline combines open datasets and Meta’s Llama ecosystem to curate massive high-quality data, which is used to train NexusRaven-V2. The data curation and training of NexusRaven-V2 do not involve any proprietary LLMs such as OpenAI’s GPT-4, which enables enterprise customers to completely own the model that is used to build copilots and agents. This is especially important for enterprise applications which require up-to-date information, quality, safety and in-depth domain customization.

NexusRaven-V2 has the potential to revolutionize workflow automation on complex software with a significantly smaller model size and higher accuracy, the company added.

Nexusflow said it is committed to thought leadership for enterprise generative AI solutions and is fully opening its model artifacts to the community to advance the enterprise adoption of generative AI using open-source models.