Jamba 1.5 Mini: benchmark results

AI21's hybrid Mamba-Transformer MoE (52B total, 12B active) with a 256K context. Provider: AI21 Labs. Released 2024-08-22. Access: API.

Unified ELO 1456 ± 1, rank #918 of 1392 rated models, from 33 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
RULER93.9Avg Accuracy (%)87.5
HELMET (128K)46.9Average Score87.1
NoLiMa92.4Base Score (%)66.7
HELM NarrativeQA74.62F1 (%)57.2
HELM NaturalQuestions (Closed)38.79F1 (%)54.4
HELM NaturalQuestions (Open)71.02F1 (%)47.8
HELM Lite45.83Mean win rate (self-reported)44.7
HELM (Stanford)41.38Mean Win Rate (%)38.9
HELM WMT 201417.9BLEU-4 (%)37.8
AA Humanity's Last Exam5.14Accuracy (%)32.7
Chatbot Arena (Text - English)1272Arena Score24.7
Chatbot Arena (Text - Creative Writing)1211Arena Score24.2

Interactive version: theaggregate.ai/model?slug=jamba-1-5-mini · How It Works · Data refreshed daily, snapshot 2026-09-05.