Llama 4 Maverick Instruct: benchmark results
Meta's natively multimodal open MoE (400B total/17B active, 128 experts) with a 1M-token context (April 2025). Provider: Meta. Released 2025-04-05. Access: Open.
Aggregate Rating 1596 ± 7, rank #456 of 1656 rated models, from 528 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CalBench - Uniform Costs - Excess Calibrated VPS | 5.3 | Excess calibrated VPS (slot-equivalent units) | 100 |
| CalBench - Varied Costs - Burden Unfairness | 0.63 | Burden unfairness (standard deviation) | 100 |
| CalBench - Varied Costs - Excess Calibrated VPS | 3.49 | Excess calibrated VPS (slot-equivalent units) | 100 |
| CalBench - Varied Costs - Excess Cost | 0.24 | Excess displacement cost (cost units) | 100 |
| MaCBench | 0.7 | Overall Score (score) | 100 |
| Open Portuguese LLM - OAB Exams | 72.85 | Accuracy (%) | 99 |
| Open Portuguese LLM - BLUEX | 81.22 | Accuracy (%) | 98.9 |
| Open Portuguese LLM - ENEM | 87.75 | Accuracy (%) | 98.8 |
| HumanAgencyBench - Avoid Value Manipulation | 65.78 | HAB dimension score (0-100, o3-judged rubric) | 97.3 |
| SEA LLM Leaderboard - SeaExam - Indonesian (Public) | 75.8 | Public Accuracy (%) | 96.7 |
| OpenCompass Knowledge - Humanities - Chinese (CompassBench 2504) | 92 | Score (%) | 96.3 |
| BenCzechMark - Czech Math Reasoning | 87.01 | Duel Win Score (%) | 94.8 |
Interactive version: theaggregate.ai/model?slug=llama-4-maverick-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.