Atlas

Models

← All models

Jamba Large 1.6

AI21 Labs · 2025-03-06 · 398.6B parameters

Jamba Large 1.6 is AI21 Labs' large-scale hybrid SSM-Transformer model released in March 2025, with approximately 398B total parameters and 94B active parameters in a MoE configuration, and a 256K-token context window. It builds on the Jamba architecture that interleaves Mamba state-space model layers with transformer attention layers to achieve faster long-context inference than comparable pure-transformer models. The model is optimized for enterprise use cases including function calling, grounded generation, structured output, and long-document retrieval.

Benchmark scores

BenchmarkScore
AIME 2024 (avg@8)0.8
AIME 2024-20250.4
AIME 2025 (avg@8)0.0
Artificial Analysis Intelligence Index5.0
Capability114.1
CorpFin v241.5
CorpFin v2 - Exact Pages50.1
CorpFin v2 - Max Fitting Context35.4
CorpFin v2 - Shared Max Context39.0
GPQA Diamond41.4
GPQA Diamond (Vals protocol average)36.1
Humanity's Last Exam (Text-Only)4.0
LegalBench72.0
LegalBench - Conclusion Tasks77.8
LegalBench - Interpretation Tasks75.4
LegalBench - Issue Tasks64.3
LegalBench - Rhetoric Tasks74.4
LegalBench - Rule Tasks67.9
LiveCodeBench (Artificial Analysis)17.2
LiveCodeBench (Vals AI public dataset)22.3
LiveCodeBench (Vals Public) - Easy57.5
LiveCodeBench (Vals Public) - Hard1.4
LiveCodeBench (Vals Public) - Medium8.1
MATH-50054.8
MATH-500 (Artificial Analysis source)58.0
MedQA Demographic Bias (Vals)50.7
MedQA Demographic Bias (Vals) - Asian50.9
MedQA Demographic Bias (Vals) - Black50.4
MedQA Demographic Bias (Vals) - Hispanic50.7
MedQA Demographic Bias (Vals) - Indigenous50.3
Loading Atlas data…