BFCL — Results
| Rank | Model | Score |
|---|---|---|
| 1 | llama-3-1-405b-instruct | 0.885 |
| 2 | llama-3-1-70b-instruct | 0.848 |
| 3 | llama-3-1-8b-instruct | 0.761 |
| 4 | nova-2-sonic | 0.745 |
| 5 | qwen3-235b-a22b | 0.708 |
| 6 | qwen3-32b | 0.703 |
| 7 | qwen3-30b-a3b | 0.691 |
| 8 | nova-pro | 0.684 |
| 9 | nova-lite | 0.666 |
| 10 | qwq-32b | 0.664 |
| 11 | nova-micro | 0.562 |
BFCL
1 phaseActive
Comprehensive executable function-call evaluation for LLMs covering serial, parallel, and multi-turn calls across Python, Java, JavaScript, and REST APIs.
Overview
BFCL
Quick answer: BFCL (Berkeley Function Calling Leaderboard) is the first comprehensive and executable function call evaluation for LLMs, covering over 2,000 question-function-answer pairs across multiple programming languages using Abstract Syntax Tree (AST) evaluation.
What is BFCL?
The Berkeley Function Calling Leaderboard (BFCL) evaluates serial and parallel function calls across Python, Java, JavaScript, and REST APIs. It covers complex use cases including multiple function calls, parallel function calls, and multi-turn interactions, with AST-based evaluation for executable accuracy.
Benchmark Details
| Property | Value |
|---|---|
| Tasks | 2,000+ question-function-answer pairs |
| Metric | Overall accuracy (AST + executable) |
| Categories | Tool calling, Reasoning |
| Modality | Text |
| Languages | Python, Java, JavaScript, REST |
Source: Yan et al. 2024. Last updated 2026-07-24.