// BENCHMARK

BFCL v3 benchmark

AI model leaderboard for the BFCL v3 benchmark. Compare how large language models score on BFCL v3, see the full ranking, and understand what this AI benchmark measures. GLM 4.5 currently leads with 77.8%. Berkeley Function Calling Leaderboard v3 — agentic tool-use benchmark across single, parallel, and multi-step calls.

Leaderboard

# Model Organization Score Variant Source
#1 GLM 4.5 Zhipu AI 77.8% official ↗
#2 GLM-4.5-Air Zhipu AI 76.4% official ↗
#3 Claude Sonnet 4 Anthropic 75.2% official ↗
#4 Claude Opus 4 Anthropic 74.4% official ↗
#5 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 73.53% thinking-cited-ring-1t official ↗
#6 o3 OpenAI 72.4% official ↗
#7 MAI-Thinking-1 Microsoft 72% official ↗
#8 Llama Nemotron Super 49B NVIDIA 71.75% reasoning-on-pass@1-avg@2 official ↗
#9 Kimi K2 Moonshot 71.1% official ↗
#10 Kimi K2 Moonshot 71.05% instruct-0905-cited-ling-1t official ↗
#11 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 70.8% reasoning official ↗
#12 Qwen3-235B-A22B Qwen 70.8% thinking official ↗
#13 Qwen3-14B Qwen 70.4% thinking official ↗
#14 Qwen3-32B Qwen 70.3% reasoning official ↗
#15 Qwen3-32B Qwen 70.3% thinking official ↗
#16 Ling-1T Ant Group 69.64% function-call official ↗
#17 Qwen3-30B-A3B Qwen 69.1% thinking official ↗
#18 GPT-4.1 OpenAI 68.9% official ↗
#19 Ring-1T Ant Group 68.82% thinking official ↗
#20 IBM Granite 4.1 IBM 68.27%
#21 Qwen3-8B Qwen 68.1% thinking official ↗
#22 Qwen3-235B-A22B Qwen 68% non-thinking official ↗
#23 o4-mini OpenAI 67.2% official ↗
#24 NVIDIA-Nemotron-Nano-12B-v2 NVIDIA 66.98% from-hf-readme official ↗
#25 NVIDIA-Nemotron-Nano-9B-v2 NVIDIA 66.9% from-hf-readme official ↗
#26 Grok 4 xAI 66.2% official ↗
#27 Qwen3-4B Qwen 65.9% thinking official ↗
#28 DeepSeek-R1-0528 DeepSeek 64.7% reasoning official ↗
#29 Granite-4.0-H-Small IBM 64.69% official ↗
#30 LFM2.5-8B-A1B Liquid AI 64.36% official ↗
#31 EXAONE 4.0 (32B) LG AI Research 63.9% reasoning official ↗
#32 DeepSeek-V3-0324 DeepSeek 63.8% official ↗
#33 gemini-2.5-pro Google DeepMind 63.31% lowthink-cited-ling-1t official ↗
#34 Qwen3-32B Qwen 63% non-thinking official ↗
#35 DeepSeek-V3.1-Terminus DeepSeek 62.01% thinking-cited-ring-1t official ↗
#36 Qwen3-14B Qwen 61.5% non-thinking official ↗
#37 gemini-2.5-pro Google DeepMind 61.36% cited-ring-1t official ↗
#38 gemini-2.5-pro Google DeepMind 61.2% official ↗
#39 Qwen3-8B Qwen 60.2% non-thinking official ↗
#40 Qwen3-30B-A3B Qwen 58.6% non-thinking official ↗
#41 Granite-4.0-H-Tiny IBM 57.65% official ↗
#42 Qwen3-4B Qwen 57.6% non-thinking official ↗
#43 Granite-4.0-H-Micro IBM 57.56% official ↗
#44 GPT-5 OpenAI 57.21% thinking-high-cited-ring-1t official ↗
#45 Qwen3-1.7B Qwen 56.6% thinking official ↗
#46 EXAONE 4.0 (1.2B) LG AI Research 52.9% reasoning official ↗
#47 Llama 4 Maverick Meta AI 52.9% official ↗
#48 DeepSeek-V3.1-Terminus DeepSeek 52.67% non-thinking-cited-ling-1t official ↗
#49 Qwen3-1.7B Qwen 52.2% non-thinking official ↗
#50 Qwen3-1.7B Qwen 52.2% non-reasoning official ↗
#51 GPT-5 OpenAI 50.27% main-cited-ling-1t official ↗
#52 Qwen3-0.6B Qwen 46.4% thinking official ↗
#53 Qwen3-0.6B Qwen 44.1% non-reasoning official ↗
#54 Qwen3-0.6B Qwen 44.1% non-thinking official ↗

Frequently asked questions about BFCL v3

What is the BFCL v3 benchmark?

Berkeley Function Calling Leaderboard v3 — agentic tool-use benchmark across single, parallel, and multi-step calls.

How is the BFCL v3 benchmark scored?

BFCL v3 is scored using the accuracy (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on BFCL v3?

As of the latest reported scores on GenAIList, GLM 4.5 achieves the highest result on BFCL v3 with a score of 77.8%.

Is a higher BFCL v3 score better?

Yes. On BFCL v3 a higher score indicates better performance, so models near the top of the leaderboard are the strongest.