BFCL-v3 — Results
| Rank | Model | Score |
|---|---|---|
| 1 | glm-4-5 | 0.778 |
| 2 | glm-4-5-air | 0.764 |
| 3 | longcat-flash-thinking | 0.744 |
| 4 | mai-thinking-1 | 0.72 |
| 5 | qwen3-next-80b-a3b-thinking | 0.72 |
| 6 | qwen3-235b-a22b-thinking-2507 | 0.719 |
| 7 | qwen3-vl-32b-thinking | 0.717 |
| 8 | qwen3-235b-a22b-instruct-2507 | 0.709 |
| 9 | qwen3-next-80b-a3b-instruct | 0.703 |
| 10 | qwen3-vl-32b-instruct | 0.702 |
| 11 | qwen3-vl-30b-a3b-thinking | 0.686 |
| 12 | qwen3-vl-235b-a22b-instruct | 0.677 |
| 13 | qwen3-vl-4b-thinking | 0.673 |
| 14 | qwen3-vl-8b-instruct | 0.663 |
| 15 | qwen3-vl-30b-a3b-instruct | 0.663 |
| 16 | qwen3-vl-4b-instruct | 0.633 |
| 17 | qwen3-vl-8b-thinking | 0.63 |
| 18 | kimi-k2-instruct | 0.591 |
BFCL-v3
1 phaseActive
Berkeley Function Calling Leaderboard v3 evaluates multi-turn tool use accuracy with AST-based verification across real-world API scenarios.
Overview
BFCL-v3
Quick answer: BFCL-v3 (Patil et al., ICML 2025) is the third major version of the Berkeley Function Calling Leaderboard, extending prior versions with multi-turn interaction evaluation to test whether LLMs can accurately use external tools across complex, stateful dialogue. Scores are computed using AST-based verification rather than LLM-as-judge, making results fully deterministic and reproducible.
At a Glance
What it tests: Single-turn and multi-turn function calling accuracy using real-world API schemas, including simple, parallel, nested, and irrelevance-detection scenarios.
Why it matters: Accurate function calling is a prerequisite for reliable AI agents; BFCL is one of the most widely-used evaluations for comparing models' ability to generate syntactically and semantically correct API calls.
Known limitations: The benchmark primarily measures format correctness via AST matching rather than end-to-end execution success; performance on BFCL may not fully reflect real-world agent reliability in complex pipelines.
What BFCL-v3 Measures
BFCL-v3 evaluates an LLM's ability to select the right API from a given function schema and generate a correctly-structured call, extended across multi-turn conversations where the model must maintain context and handle follow-up requests. The benchmark covers multiple difficulty levels: simple single-function calls, parallel calls, function composition (nested), and scenarios where no applicable function exists (irrelevance detection).
Unlike LLM-as-judge systems, BFCL uses Abstract Syntax Tree (AST) evaluation — the generated function call is parsed and compared to the ground truth at the structural level — which eliminates subjectivity and makes scores fully reproducible regardless of which judge model is used.
Version 3 expanded on BFCL-v2 by adding multi-turn interaction scenarios, reflecting the reality that agentic applications require sustained tool use over multiple conversation turns, not just isolated single-shot function calls. The benchmark also includes both FC-native (models with dedicated tool-calling APIs) and Prompt-based (text-generation fallback) evaluation tracks.
Benchmark Specifications
| Field | Value |
|---|---|
| Task category | Tool use / Agent |
| Metric | Overall accuracy (unweighted average across sub-categories) |
| Number of tasks | ~2,251 across all categories |
| Saturation | Low |
| Created by | Patil et al. (UC Berkeley) |
| Source paper | Patil et al. ICML 2025 |
| GitHub | ShishirPatil/gorilla |
| Dataset | gorilla-llm/Berkeley-Function-Calling-Leaderboard |
How BFCL-v3 Is Scored
The primary score is the unweighted average accuracy across all evaluation sub-categories (simple, parallel, nested, multiple, irrelevance, and multi-turn). For each test case, the model's output is parsed as an AST and compared against the ground truth function call; a match is awarded 1 point and a mismatch 0. Models are evaluated in two modes: FC (native function-calling API) and Prompt (structured generation via system prompt), with results reported separately where applicable.
FAQ
What is BFCL-v3? BFCL-v3 is the third version of the Berkeley Function Calling Leaderboard, an automated benchmark for evaluating how accurately LLMs can invoke external tools and APIs, extended to include multi-turn interaction scenarios.
How is BFCL-v3 scored? Each model output is parsed as an Abstract Syntax Tree (AST) and compared against a ground-truth function call. The overall score is the unweighted average accuracy across all sub-categories including simple, parallel, nested, and multi-turn scenarios.
What is a good BFCL-v3 score? Scores above 70% represent strong function-calling capability. As of mid-2026, the top models (GLM-4.5, GLM-4.5-Air, LongCat-Flash-Thinking) score between 74–78%.