Jamba Reasoning 3B — benchmark results
AI21's Apache-2.0 3B reasoning model (October 2025) using the hybrid SSM-Transformer Jamba architecture, running 250K+ token contexts on-device. Provider: AI21 Labs. Released 2025-10-08. Access: Open.
Unified ELO 1399 ± 30, rank #1230 of 1776 rated models, from 36 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA IFBench | 52.45 | Accuracy (%) | 62.7 |
| CritPt | 0 | Accuracy (self-reported) | 30.3 |
| AA Omniscience - Software Engineering (SWE) - PHP | 16 | Accuracy (%) | 28.8 |
| AA CritPt | 0 | Accuracy (%) | 27.1 |
| AA-LCR | 7 | Score (self-reported) | 25.8 |
| AA Humanity's Last Exam | 4.59 | Accuracy (%) | 24.1 |
| AA Omniscience - Software Engineering (SWE) - Julia | 4 | Accuracy (%) | 23.7 |
| AA LiveCodeBench | 20.95 | Pass@1 (%) | 22.1 |
| AA MMLU-Pro | 57.75 | Accuracy (%) | 20.6 |
| AA Long Context Reasoning | 7 | Accuracy (%) | 19.5 |
| AA Omniscience - Software Engineering (SWE) - Swift | 24 | Accuracy (%) | 19.5 |
| AA Omniscience - Software Engineering (SWE) - Go | 8 | Accuracy (%) | 17.5 |
Interactive version: theaggregate.ai/model?slug=jamba-reasoning-3b · How the rankings work · Data refreshed daily, snapshot 2026-07-22.