Gorilla OpenFunctions v2

Gorilla OpenFunctions v2 is an open-source function-calling model from UC Berkeley's Gorilla project, released February 26, 2024 alongside the Berkeley Function-Calling Leaderboard (BFCL). It extends chat-completion-style models to formulate executable API calls from natural-language instructions, adding multiple function selection, parallel function calls, combined multiple-and-parallel calls in a single generation, relevance detection (chat vs. call), and support for Python, Java, JavaScript, and native REST parameter types. On the BFCL it scored 83.55–83.67% overall accuracy — essentially on par with GPT-4-0125-Preview's 83.8% and well ahead of GPT-3.5 and Mistral Medium — making it the state of the art for open-source models at release. It is licensed Apache 2.0 and hosted under the gorilla-llm Hugging Face org.

Limitations: v2 is a 2024-era model and the Gorilla team has since moved toward v3 with more languages and multi-turn capability; check the leaderboard for current rankings before choosing it for new work. It is a function-calling specialist rather than a general assistant, and like all tool-calling models it needs a runtime (such as the Gorilla Execution Engine / GoEx) with safety guarantees for executing generated actions.