// the find
MeetKai/functionary
Chat language model that can use tools and interpret the results
Functionary is a fine-tuned LLM series built specifically for function calling, released as open weights on Hugging Face. It exposes an OpenAI-compatible API so you can drop it in as a self-hosted replacement for GPT-4's tool-use capabilities. The repo is explicitly deprecated as of the README — what's here is a historical snapshot.
The prompt template approach is technically honest: function schemas are converted to TypeScript-style type definitions rather than raw JSON, which gives the model a more natural token space to reason about. The ToolSandbox benchmark results (68.87 for medium-v3.1, competitive with Claude-3-Opus) are more meaningful than Berkeley leaderboard numbers because ToolSandbox tests stateful multi-turn tool use, not just single-call extraction. GGUF variants are provided for every model version, so llama.cpp inference is viable without owning a rack. The training code ships in the repo, including DeepSpeed and FSDP configs, so you can actually fine-tune on your own data rather than being stuck with what they released.
The repo is abandoned — the README says so in a warning block at the top, and the last meaningful model release was December 2024. Adopting this for anything production-facing means you own it from day one. The medium models (v3.x) require 4×A6000 or 2×A100 80GB — that's not a hobbyist setup, and the small models trade enough accuracy that the benchmark advantage over alternatives narrows considerably. The llama.cpp integration is explicitly noted as lagging behind, with v2.5 not integrated at all, so the 'works with llama-cpp-python' story is incomplete. There's no guidance on what the team is doing instead, so if you're evaluating this today you're flying blind on the roadmap.