DeepSeek V2: benchmark results

DeepSeek's 236B open MoE (21B active) with Multi-head Latent Attention and a 128K context. Provider: DeepSeek. Released 2024-05-06. Access: Open.

Unified ELO 1537 ± 1, rank #504 of 1392 rated models, from 9 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
ARC Challenge (AI2)92.2Accuracy (%)96.1
WinoGrande86.3Accuracy (%)92.5
HellaSwag87.1Accuracy (%)92.1
Big-Bench Hard78.8Average (%)87.5
PIQA83.9Accuracy (%)86.7
MMLU78.4Accuracy (%)77
MixEval51.7Score68.6
TriviaQA80Accuracy (%)64.1
BenchGecko Score86.1Composite Score42.1

Interactive version: theaggregate.ai/model?slug=deepseek-v2 · How It Works · Data refreshed daily, snapshot 2026-09-05.