Reka Flash 3 — benchmark results

Reka's open-weights 21B reasoning model trained from scratch (Apache 2.0, 32K context), aimed at low-latency and on-device use (March 2025). Provider: Reka. Released 2025-03-10. Access: Open.

Unified ELO 1529 ± 14, rank #688 of 1776 rated models, from 152 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
EuroEval German NLU - Sb10K59.6Sentiment classification Score (%)93.1
EuroEval French NLU - Eltec71.29Named entity recognition Score (%)92.8
EuroEval Danish NLU - Angry Tweets57.97Sentiment classification Score (%)92.2
EuroEval English NLU - SST-569.19Sentiment classification Score (%)91.9
EuroEval Italian NLU - Sentipolc1665.9Sentiment classification Score (%)90.5
EuroEval Spanish NLU - CoNLL ES75.76Named entity recognition Score (%)89.8
EuroEval English NLU - CoNLL EN81.67Named entity recognition Score (%)88
EuroEval German NLU - GermEval68.07Named entity recognition Score (%)82.9
EuroEval Portuguese NLU - MultiWikiQA PT74.48Reading comprehension Score (%)82
EuroEval Dutch67Average Score (%)81.2
EuroEval Portuguese NLU - HAREM51.85Named entity recognition Score (%)81.2
EuroEval Dutch NLU71.73NLU Average Score (%)80.8

Interactive version: theaggregate.ai/model?slug=reka-flash-3 · How the rankings work · Data refreshed daily, snapshot 2026-07-22.