Nejumi 4 - BFCL - Relevance Detection: leaderboard
Berkeley Function Calling Leaderboard as adapted and sampled by Nejumi Leaderboard 4 (Weights & Biases Japan): accuracy at calling a function when a relevant one is available (live relevance category). Accuracy (%), higher is better; the overall accuracy is Nejumi 4 - GLP - Function Calling.
Metric: Accuracy (%). Source: nejumi.ai. Saturation forecast: Not forecast. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick Instruct FP8 | 100 | #424 |
| 2 | Grok 4.5 (High) | 100 | #54 (Grok 4.5) |
| 3 | Grok 4.20 0309 (Reasoning) | 100 | #120 (Grok 4.20 0309) |
| 4 | qwen2.5-bakeneko-32B-instruct-v2 | 100 | #570 |
| 5 | ELYZA-Shortcut-1.0-Qwen-32B | 100 | #560 |
| 6 | Gemma 3 12B (IT) | 94.44 | #672 |
| 7 | Qwen 3 Next 80B A3B Instruct | 94.44 | #381 |
| 8 | Qwen 3 235B A22B 2507 (Thinking) | 94.44 | #263 |
| 9 | Llama 3.1 405B Instruct FP8 | 94.44 | #400 |
| 10 | Llama-3.3-Swallow-70B-Instruct-v0.4 | 94.44 | #540 |
| 11 | Grok 4 | 88.89 | #189 |
| 12 | GPT-4o Mini (2024-07-18) | 88.89 | #568 |
| 13 | DeepSeek V3.2 (Thinking) | 88.89 | #210 (DeepSeek V3.2) |
| 14 | Qwen 3 30B A3B 2507 (Thinking) | 88.89 | #379 |
| 15 | DeepSeek V4 Pro (Max) | 88.89 | #112 (DeepSeek V4 Pro) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nejumi-4-bfcl-relevance-detection · How It Works · Data refreshed daily, snapshot 2026-11-05.