Berkeley Function Calling Leaderboard (BFCL)
Confidence 0.85 · 3 sources · last confirmed 2026-09-04
The de-facto standard benchmark for LLM function calling / tool use, built and maintained by the Gorilla project at UC Berkeley’s Sky Computing Lab. Live at gorilla.cs.berkeley.edu/leaderboard.html; evaluation code open on GitHub; refreshed roughly quarterly, tracking on the order of 110 model families with vendor-supplied runs plus verification.
On the ai-benchmarks roster it is the one domain where the field has converged on a single vendor-neutral instrument. Full treatment on the source page: 2025-07-13-patil-berkeley-function-calling-leaderboard.
What it measures
5,551 question–function–answer pairs across Python, Java, JavaScript, REST and SQL, in four sections: single-turn (simple / multiple / parallel / parallel-multiple / irrelevance), crowd-sourced (2,251 entries curated from >67,000 real community-contributed queries, 15+ languages), multi-turn (eight API suites, 1,000 queries, with missing-parameter, missing-function and long-context variants), and agentic (web search, memory, SQL).
Its methodological signature is Abstract Syntax Tree sub-string matching as a proxy for executing the call — validated against execution on a subset, which is what makes the shortcut defensible. Multi-turn entries are scored by both final-state match and required-call-path match.
Version history: v1 introduced AST evaluation, v2 added enterprise and community-contributed functions, v3 added multi-turn, v4 added holistic agentic evaluation.
Two things to know before citing a BFCL number
- “Overall Accuracy” is a composite index, not a per-call success rate. It weights sections of wildly different difficulty. A model at 66 overall can be at 95 on single-turn calls and at 6 on memory. Stated cleanly by Sokolenko answering exactly this audience question: “It’s a combined score of four other scores. Think of it as an index… accuracy is the wrong name for it.”
- Every rank is an artefact of its date. xLAM-2-70b and -32b held ranks 1 and 2 in April 2025 (APIGen-MT, Table 1); by the December 2025 snapshot the top was Claude Opus 4.5 / Gemini 3 / GLM 4.6 and xLAM-2 sat around position 18. The models did not regress; the board moved.
The structural finding
Single-turn function calling is saturated — frontier models cluster in the high 80s and 90s. Multi-turn state, memory and long-horizon reasoning are not: the best memory score in the paper’s Table 1 is 12.0 out of 100. Any claim of agentic parity between two models that rests on single-turn numbers is resting on the saturated half.
Appears in this wiki via
- 2025-07-13-patil-berkeley-function-calling-leaderboard — the ICML 2025 paper (primary).
- 2025-04-04-prabhakar-salesforce-apigen-mt-xlam-2 — xLAM-2 evaluated on BFCL v3; ranks 1, 2 and 4 as of 3 April 2025.
- 2026-08-25-sokolenko-pycon-de-demystifying-agentic-ai-small-language-models — used on stage as the shopping list for a locally-runnable tool-calling model; the composite-index caveat.
- 2025-06-02-belcak-nvidia-small-language-models-future-agentic-ai — argues (barrier B2) that measuring SLMs on agentic benchmarks rather than generalist ones is what reveals their sufficiency; BFCL is the instrument that makes this possible.