Report Abuse

Basic Information

Functionary is an open source project that provides a language model and associated server tooling designed to interpret and execute external functions or plugins. The system converts function definitions into typed prompts and injects them as system messages so the model can decide when and how to call functions, whether serially or in parallel, and how to use their outputs to produce grounded responses. It offers model artifacts targeted at different resource envelopes and capabilities, including models with code-interpreter features and very long context lengths. The repository supplies server implementations that are compatible with vLLM, SGLang, Text-Generation-Inference and llama-cpp, an OpenAI-compatible chat API surface, examples for calling real Python functions, and deployment options such as Docker and serverless Modal. Documentation and examples show how to provide tools as JSON Schema-like definitions and how to make raw or OpenAI-style requests to the running service.

Links

Categorization

App Details

Features
Functionary centers on robust function-calling support and deployable inference servers. It supports single and parallel function calls and multi-turn interactions with the ability to follow up on missing arguments. Models include variants with code-interpreter capability and very long context windows. Server options provided include vLLM and SGLang servers, a TGI wrapper for text-generation-inference, llama-cpp integration and Dockerized deployment. The project implements dynamic LoRA adapter loading for vLLM, OpenAI-compatible chat endpoints, examples for calling real Python functions via chatlab, and serverless deployment through Modal. The README documents available model builds and VRAM footprints, provides example request formats for OpenAI-compatible and raw HTTP clients, and reports evaluation results on benchmarks showing high function-calling accuracy relative to other open-source and proprietary models.
Use Cases
Functionary helps developers and teams integrate language models that can reliably call external tools and services as part of conversational workflows. By accepting function definitions as JSON Schema-like objects and exposing OpenAI-compatible endpoints, it simplifies wiring a model to existing APIs, Python functions and toolchains without custom parsing logic. The provided servers and deployment recipes reduce operational friction for local, containerized or serverless hosting, and LoRA support enables runtime adapter experimentation. Built-in examples and use cases illustrate practical tasks such as trip planning, property valuation and customer complaint parsing, which demonstrate how function calls can produce structured inputs for downstream systems. Evaluation results and model variants give practitioners guidance about capability and resource needs, making it suitable for prototyping grounded agent behaviors and production tool-calling features.

Please fill the required fields*