Berkeley Function-Calling Leaderboard
UC BerkeleyThe Berkeley Function-Calling Leaderboard (BFCL) measures how accurately language models call functions and tools. Its evaluation package generates responses from hosted models through their APIs or from local models served with vLLM or SGLang, then grades them against expected calls, with categories that grew from single, parallel and multiple calls to multi-turn conversations, agentic web search, memory and prompt-format sensitivity. UC Berkeley's Gorilla team maintains it and publishes the results as a live leaderboard.
The code lives in one folder of the Gorilla repository, which also holds Gorilla's other projects, so the repository's stars are not this product's.
Openness
5 high confidence- license
- Apache-2.0 (OSI; the repository LICENSE body is the Apache 2.0 text)
- source
- public (the berkeley-function-call-leaderboard package in the Gorilla repository)
- core features withheld
- no — a university research project with nothing sold beside it, and the tree has no enterprise path
The evaluation code is Apache-2.0 and published in full, and the leaderboard's own results are produced from a named commit of it. A university team runs it with no paid tier.
- https://gorilla.cs.berkeley.edu/leaderboard.html recorded 2026-11-12
The leaderboard on Berkeley's cs.berkeley.edu site: "Checkout code and data", "Models are evaluated using commit f7cf735", and contributions by contacting the team; no pricing or paid tier.
- https://pypi.org/pypi/bfcl-eval/json recorded 2026-11-12
PyPI bfcl-eval 2026.3.23, uploaded 2026-03-23, license "Apache 2.0", summary "Berkeley Function Calling Leaderboard (BFCL)", repository github.com/ShishirPatil/gorilla/tree/main/berkeley-function-call-leaderboard.
- https://raw.githubusercontent.com/ShishirPatil/gorilla/main/LICENSE recorded 2026-11-12
The repository LICENSE body: the Apache License, Version 2.0 text.
- https://ungh.cc/repos/ShishirPatil/gorilla/files/main recorded 2026-11-12
The 534-file tree of ShishirPatil/gorilla at `main`: berkeley-function-call-leaderboard beside gorilla, openfunctions, goex, raft and agent-arena, and no ee/, enterprise/, commercial/ or proprietary/ path. Path absence only, paired with the leaderboard page cited beside it.
Adoption
3 medium confidencePyPI downloads of bfcl-eval, the evaluation package the README installs. Automated evaluation pipelines can reinstall a package often, so the figure counts installs rather than users.
- https://pypistats.org/api/packages/bfcl-eval/recent recorded 2026-11-12
last_month = 120,911 downloads for the package bfcl-eval.
- https://raw.githubusercontent.com/ShishirPatil/gorilla/main/berkeley-function-call-leaderboard/README.md recorded 2026-11-12
README: "pip install bfcl-eval # Be careful not to confuse with the unrelated `bfcl` project on PyPI!", beside an editable install from a clone.
Capability
4 medium confidenceBFCL covers tool calling in depth, from single calls to multi-turn and agentic tasks. It sits a step below SWE-bench, which runs each answer against a real repository's own tests, because most BFCL categories compare a model's calls with expected calls rather than running the resulting program.
- https://gorilla.cs.berkeley.edu/leaderboard.html recorded 2026-11-12
"The Berkeley Function Calling Leaderboard (BFCL) V4 evaluates the LLM's ability to call functions (aka tools) accurately"; "BFCL-v1 introducing AST as an evaluation metric"; "Last Updated: 2026-04-12".
- https://raw.githubusercontent.com/ShishirPatil/gorilla/main/berkeley-function-call-leaderboard/README.md recorded 2026-11-12
README: "the first comprehensive and executable function call evaluation"; blog series "BFCL v1: Simple, Parallel, and Multiple Function Call eval with AST", "BFCL v3: Multi-Turn & Multi-Step", "BFCL V4 Part 1: Agentic Web Search", "Part 2: Agentic Memory Management", "Part 3: Agentic Format Sensitivity"; vLLM and SGLang backends for locally hosted models.
Verified 2026-11-12, except some scores, still awaiting re-confirmation