Atlas

Models

← All models

Jamba Mini 1.6

AI21 Labs · 2025-03-06 · 51.6B parameters

Jamba Mini 1.6 is AI21 Labs' smaller hybrid SSM-Transformer model released in March 2025, with approximately 52B total parameters and 12B active parameters in a MoE configuration and a 256K-token context window. It uses the same Mamba-Transformer hybrid architecture as the larger Jamba models to deliver fast long-context inference with lower hardware requirements. The model targets enterprise workloads requiring low latency and accurate long-context retrieval, including function calling and structured output tasks.

Benchmark scores

BenchmarkScore
AIME 2024 (avg@8)0.8
AIME 2024-20250.4
AIME 2025 (avg@8)0.0
Artificial Analysis Intelligence Index2.6
Capability101.7
CorpFin v238.0
CorpFin v2 - Exact Pages45.0
CorpFin v2 - Max Fitting Context32.1
CorpFin v2 - Shared Max Context37.1
GPQA Diamond33.3
GPQA Diamond (Vals protocol average)32.1
Humanity's Last Exam (Text-Only)4.6
LegalBench69.7
LegalBench - Conclusion Tasks73.1
LegalBench - Interpretation Tasks69.3
LegalBench - Issue Tasks72.0
LegalBench - Rhetoric Tasks65.9
LegalBench - Rule Tasks68.3
LiveCodeBench (Artificial Analysis)7.1
LiveCodeBench (Vals AI public dataset)9.9
LiveCodeBench (Vals Public) - Easy27.6
LiveCodeBench (Vals Public) - Hard0.3
LiveCodeBench (Vals Public) - Medium1.8
MATH-50025.4
MATH-500 (Artificial Analysis source)25.7
MedQA Demographic Bias (Vals)52.5
MedQA Demographic Bias (Vals) - Asian52.5
MedQA Demographic Bias (Vals) - Black51.1
MedQA Demographic Bias (Vals) - Hispanic52.0
MedQA Demographic Bias (Vals) - Indigenous52.8
Loading Atlas data…