Gorilla

Gorilla is a UC Berkeley research project that ships an open-source LLM (gorilla-openfunctions-v2, ~6.91B params) trained specifically to translate natural-language intents into accurate API calls. Beyond the base model, the project includes GoEX (a runtime that executes LLM-generated actions with undo and damage-confinement safety primitives), RAFT (a fine-tuning recipe for retrieval-augmented generation), and the Berkeley Function-Calling Leaderboard (BFCL) which benchmarks function-calling quality across 2,000+ test cases in Python, Java, and REST.

It's aimed at developers and ML teams who want an open, self-hostable alternative to closed function-calling APIs from OpenAI or Anthropic. Everything is Apache 2.0, with weights on HuggingFace and a hosted demo plus Colab notebook for kicking the tires. Because it's research-led, the polish is uneven compared to commercial offerings, but the BFCL leaderboard and ongoing publications give it real credibility as a reference implementation for tool-using agents.

Integrations cover Python, Java, and REST out of the box, and the project pairs naturally with agent frameworks that need a local function-calling backbone. The caveat is that 'product' here means GitHub code plus model weights, not a managed SaaS, so you handle hosting, eval, and ops yourself.