Llama 3.2 90B: benchmark results

Meta's 90B vision-language base model from Llama 3.2, adding cross-attention image adapters to Llama text weights with a 128K context. Provider: Meta. Released 2024-09-25. Access: Open.

Unified ELO 1533 ± 1, rank #526 of 1392 rated models, from 22 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
Translation Set1 to en spBleu43.7Set1→en spBLEU (self-reported)75
AI for Education Pedagogy - Technology82.08Accuracy (%)69.1
MMMU Benchmark60.3Validation Score66.1
AI for Education Pedagogy - Maths76.19Accuracy (%)47.9
AI for Education Pedagogy76.31Accuracy (%)45.4
AI for Education Pedagogy - Secondary75Accuracy (%)44.3
AI for Education Pedagogy - Social studies74.55Accuracy (%)43.9
AI for Education SEND70.64Accuracy (%)41.3
AI for Education Pedagogy - Primary79.81Accuracy (%)39.7
AI for Education Pedagogy - Science74.86Accuracy (%)35.7
Translation Set1 to en COMET2288.5Set1→en COMET22 (self-reported)29.2
AI for Education Visual Maths - Geometry36.92Accuracy (%)27.7

Interactive version: theaggregate.ai/model?slug=llama-3-2-90b · How It Works · Data refreshed daily, snapshot 2026-09-05.