Gorilla is UC Berkeley’s research project on teaching large language models to call APIs and tools accurately. It introduced retriever-aware training for function calling, maintains the widely used Berkeley Function-Calling Leaderboard, and has earned roughly 13,100 stars on GitHub.
When an LLM invokes tools, a hallucinated function name or argument breaks the whole workflow. Gorilla tackled this by fine-tuning models on APIBench, a large dataset of real API calls spanning thousands of machine learning and cloud APIs, and by training the model to use a retriever at inference time so it can adapt when APIs change and hallucinate less. The project grew into an ecosystem: the APIZoo collects thousands of documented APIs for training and evaluation, the OpenFunctions models apply the approach to open weights, and the Berkeley Function-Calling Leaderboard (BFCL) has become a standard public benchmark that model providers cite when comparing tool-calling ability. The repository contains the datasets, training recipes and evaluation harness, making it as much shared infrastructure for the agent research community as a single model release. It remains actively maintained through 2026.
Gorilla is free and open source under the Apache-2.0 license. The models, datasets, leaderboard and evaluation code are publicly available for research and commercial use; you only supply your own compute and any model API keys.
Gorilla is for researchers and engineers building tool-using agents, evaluating how reliably different models call functions, or training their own models for API use, and for anyone who wants a rigorous public benchmark of LLM tool-calling ability.