Atlas

Models

← All models

Jamba 1.5 Mini

AI21 Labs · 2024-08-22 · 51.6B parameters

Jamba 1.5 Mini is AI21 Labs' compact hybrid SSM-Transformer model released in August 2024, with approximately 52B total parameters and 12B active parameters and a 256K-token context window. It shares the same Mamba-Transformer hybrid architecture as Jamba 1.5 Large, offering faster long-context inference at a smaller scale suitable for lower-resource deployments. The model is intended for enterprise chat, retrieval, and instruction-following tasks where low latency is a priority.

Benchmark scores

BenchmarkScore
Artificial Analysis Intelligence Index2.7
Capability106.0
CorpFin v233.9
CorpFin v2 - Exact Pages39.6
CorpFin v2 - Max Fitting Context30.1
CorpFin v2 - Shared Max Context31.9
GPQA Diamond30.2
Humanity's Last Exam (Text-Only)5.1
LegalBench66.6
LegalBench - Conclusion Tasks68.9
LegalBench - Interpretation Tasks69.2
LegalBench - Issue Tasks61.2
LegalBench - Rhetoric Tasks64.7
LegalBench - Rule Tasks69.1
LiveCodeBench (Artificial Analysis)6.2
MATH-500 (Artificial Analysis source)35.7
MedQA Demographic Bias (Vals)55.2
MedQA Demographic Bias (Vals) - Asian55.3
MedQA Demographic Bias (Vals) - Black54.4
MedQA Demographic Bias (Vals) - Hispanic55.0
MedQA Demographic Bias (Vals) - Indigenous54.0
MedQA Demographic Bias (Vals) - Unbiased56.5
MedQA Demographic Bias (Vals) - White55.9
MMLU69.9
MMLU Pro37.1
SciCode (Artificial Analysis)8.0
TaxEval v241.9
TaxEval v2 - Correctness29.8
TaxEval v2 - Stepwise Reasoning54.0
Loading Atlas data…